Residual network-based general-inductance integrated waveform super-resolution sensing method

Abstract

The invention relates to a residual network-based communication and sensing integrated waveform super-resolution sensing method. The method comprises the following steps of: 1, constructing a network model; step 2, signal processing flow; and step 3, a model training process. The method has the advantages that the good image processing capacity of the residual network and the feature selection capacity of an attention mechanism are utilized, a two-dimensional MUSIC perception spectrum image is processed, and higher precision is achieved on separation of similar targets; according to the method, the estimation problem of the sensing parameters is modeled into a regression prediction problem, the sensing parameters are directly output, and compared with a multi-classification indirect estimation method which divides grids and predicts spectrum peak positions, the sensing performance of the method is not limited by the grid resolution.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a Doppler and delay estimation method for a receiving end of an orthogonal frequency division multiplexing system in a communication perception integration scenario. Background Art

[0002] Integrated sensing and communication (ISAC) is one of the six application scenarios of the sixth generation mobile network (6G) and is also a key technology for realizing the deep integration of the physical world and the digital world. 6G has an increasingly urgent demand for intelligent systems with both communication and perception functions. In this context, communication and perception integration has become a key enabling technology to promote the integration and coordination of communication and perception functions and improve the spectrum utilization efficiency of the system, attracting great interest and widespread attention from the academic community. In 6G, integrated waveform design is the core technology of ISAC design. Adding perception functions to the existing OFDM (Orthogonal Frequency Division Multiplexing, OFDM) mobile communication system without making major changes to the cellular system is a relatively direct and low-cost way. Systems using OFDM waveforms have high resolution and strong anti-multipath fading capabilities, showing excellent performance. In order to achieve and optimize the perception performance of OFDM systems, many traditional radar perception algorithms have been tried to be applied to OFDM systems. The traditional OFDM speed-distance joint perception algorithm can be traced back to 2D-DFT, as well as super-resolution algorithms such as MUSIC and Capon, which are also used in OFDM systems. However, the accuracy that can be achieved by traditional super-resolution algorithms is limited by the time width and bandwidth of OFDM signals. In a multi-target scenario, when the distance difference or speed difference between two targets is less than the perception resolution, the corresponding perception spectrum peaks of the two targets on the perception spectrum will interfere with each other. As the perception parameters gradually approach, the two spectrum peaks gradually merge until the traditional "super-resolution" algorithm cannot separate them.

[0003] In recent years, more and more studies have focused on the application of machine learning methods in super-resolution estimation. Traditional multi-layer perceptrons have been tried to be applied to the estimation of the angle of arrival of a single snapshot, achieving low-complexity estimation of multiple targets. In the field of autonomous driving, the problem of two-dimensional angle of arrival estimation of a car has been shown to be expressed as a block sparse recovery problem, and the problem is solved by a sparse Bayesian learning network to achieve the estimation of perception parameters. The channel estimation problem in large-scale multi-input and multi-output systems can also be solved to a certain extent by introducing deep learning networks. Although these studies have demonstrated the potential of machine learning methods in the problem of angle of arrival estimation and can also be applied to OFDM perception estimation, their performance and complexity are difficult to balance. Therefore, a perception estimation method that achieves an effective compromise between performance and complexity remains to be discovered. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention proposes a synaesthesia integrated waveform super-resolution perception method based on residual network.

[0005] The residual network-based synaesthesia integrated waveform super-resolution perception method includes:

[0006] Step 1: Network model construction

[0007] Step 1.1: The first layer of the neural network consists of a normal convolutional layer and a pooling layer. The convolutional layer has a kernel size of 3×3, a stride of 1, and 64 output channels. A batch normalization layer is added after the convolutional layer, and the ReLu function is used as the output function of the batch normalization layer. The maximum pooling layer is placed after the ReLu layer, with a kernel size of 3×3 and a stride of 2.

[0008] Step 1.2: The second and third layers of the neural network are each composed of a classic residual block. Each residual block contains two convolutional layers. The convolution kernel size of each convolutional layer is set to 3×3, the step size is 1, the output channel of the second residual block is 64, and the output channel of the third residual block is 128;

[0009] Step 1.3: The fourth and fifth neural network layers add an attention mechanism based on the residual block. The internal parameter settings of the fourth and fifth residual blocks are the same as the first two layers. The number of output channels is set to 256 and 512 respectively. A layer of SENet (Squeeze-and-excitation net) is added after the residual block, and the dimensionality reduction ratio is set to 8 and 16 respectively;

[0010] Step 1.4: After the fifth layer, a global average pooling layer is used to reduce the dimension of the data, followed by a fully connected layer as the output layer. The number of output neurons is set to in is the number of perceived targets;

[0011] Step 2: Signal processing flow

[0012] Step 2.1: Set the parameters N, M and Where: N is the number of OFDM symbols, M is the number of subcarriers, is the modulation order, and the source generates a length of The serial bit data stream is processed The digital modulation of the order is mapped to MN constellation symbols, where: m = 0, ..., M-1, n = 0, ..., N-1;

[0013] Step 2.2: The MN constellation symbols are first placed on the time-frequency domain grid through serial-to-parallel transformation to obtain the symbols to be transmitted in the time-frequency domain, which are recorded as A[m, n]. The transmitter performs M-point discrete inverse Fourier transform (IDFT) on each row of A[m, n] to obtain discrete time domain samples, and then converts them into continuous time domain signals to be transmitted x(t) through the transmission pulse shaping function, as shown in the following formula (1):

[0014]

[0015] In formula 1, Δf is the subcarrier spacing, T s is the duration of an OFDM symbol with a cyclic prefix (CP), satisfying T s =T+T cp , and T = 1 / Δf is the duration of the OFDM symbol information part, T cp is the duration of the CP part, g tx (t) is the transmission pulse shaping function;

[0016] Step 2.3: Set perception parameters ρ k , d k and v k ,in is the number of perceived targets, ρ k , d k and v k are the channel gain, distance and speed corresponding to the kth target respectively, and there is

[0017] Step 2.4: Construct the time-delay domain channel impulse response based on the perception parameters where δ(·) is the Dirac delta function, τ k and v kare the time delay and Doppler corresponding to the kth path, satisfying τ k =2d k / c and v k =2v k f c / c, c represents the speed of light, f c Indicates the center carrier frequency of the transmitted signal;

[0018] Step 2.5: The transmitted signal reaches the sensing receiver through the sensing channel. The received signal r(t) is obtained by integrating the transmitted signal x(t) and the time-delay domain channel impulse response h(t, τ), as shown in the following equation (2):

[0019]

[0020] In formula (2), n(t) is the additive white Gaussian noise received by the receiver. The receiving end can obtain a discrete received signal by sampling the received signal at a time interval of T / M, removing the part of the sampled data that belongs to the cyclic prefix, and performing DFT on every M sampling points, that is, obtaining the received signal matrix Y[m, n] in the time-frequency domain: as shown in formula (3)

[0021]

[0022] The time-frequency domain receiving matrix is ​​regarded as the result of multiplying the transmitted symbol A[m,n] by the channel complex gain plus the noise, then Y=A⊙H(d,v)+N, where ⊙ represents the Hadamard product, H(d,v) is the channel impulse response matrix in the time-frequency domain, and Represents the perceptual parameter vector to be estimated, let represents the channel gain vector, then H(d, v) is further decomposed into H(d, v) = U(d)diag(Γ)B(v) T , where U(d) and B(v) represent the phase shift matrices corresponding to the delay and Doppler, respectively, as shown in equation (4):

[0023]

[0024] In formula (4), Step 2.6: The estimation algorithm needs to receive channel-related information in the signal matrix. The received matrix Y needs to be divided by the transmitted symbol A by element-wise division to obtain the estimated matrix O:

[0025]

[0026] In the above formula (5), the row direction of the estimated matrix O contains the distance information of each target, and the column direction contains the speed information of each target. represents the noise matrix;

[0027] Step 2.7: Rearrange O so that the column vectors are joined together one by one, placing the distance information and speed information on the same dimension, and defining the turning vector in: represents the Kronecker product, then the rearranged O is expressed as follows:

[0028]

[0029] In formula (6), represents the rearranged vector, and vec(·) represents the column vectorization operation;

[0030] Step 2.8: Transpose q with its conjugate H Multiply and calculate the covariance matrix Q xx =qq H ;

[0031] Step 2.9: Calculate the covariance matrix Q xx The eigenvalues ​​and corresponding eigenvectors of Q xx The eigenvalue decomposition Q xx =EΛE H , where: Λ=diag(λ1,...,λ MN ), E = [e1, ..., e MN ],λ l and e l They are the covariance matrices Q xx The eigenvalues ​​and corresponding eigenvectors of

[0032] Step 2.10: Sort the calculated eigenvalues ​​in descending order so that the sorted eigenvalue sequence satisfies in The eigenvectors are also rearranged in the order of eigenvalues, and the rearranged eigenvector matrix is ​​recorded as

[0033] Step 2.11: Maximum front The eigenvalues ​​are considered The eigenvalues ​​corresponding to the reflected signals of the targets, and the rest The eigenvalues ​​are considered to be noise eigenvalues. According to the covariance matrix Q xx The received signal is decomposed into signal subspace and noise subspace according to the eigenvalue decomposition and sorting results. The result of subspace decomposition is expressed as follows:

[0034]

[0035] In formula (7), and Represent the signal subspace and the noise subspace respectively. Based on the above decomposition operation, the two-dimensional MUSIC spectrum is expressed as follows (8):

[0036]

[0037] In formula (8), d and v represent distance and speed respectively, c(d, v) is the steering vector defined in step 2.7, represents the noise subspace, (·) H represents the conjugate transpose.

[0038] Step 2.12: Set parameter N d and N v , respectively represent the number of sampling points when sampling the two-dimensional MUSIC spectrum along the distance axis and the velocity axis;

[0039] Step 2.13: Based on the steering vectors u(d) and b(v) described in step B5, calculate the maximum distance range D estimated by the two-dimensional MUSIC algorithm range and maximum speed range V range , respectively satisfying and That is, we have the following formula (9):

[0040]

[0041] Step 2.14: Determine the maximum distance d that the target can reach in the algorithm application scenario max and the maximum rate v max , satisfying 0≤d k ≤d max and -v max ≤v k ≤v max ,in: At the same time, it is necessary to ensure that the waveform parameters are set so that D range ≥d max and V range ≥v max / 2;

[0042] Step 2.15: According to the parameters set or calculated in the above three steps, the two-dimensional MUSIC spectrum is sampled along the distance axis and the velocity axis. The sampled two-dimensional discrete MUSIC spectrum is recorded as P music , as shown in formula (10):

[0043]

[0044] In formula (10), η d =0, ..., N d , n v =0, ..., Nv The usual practice is to set each sampling point to D range n d / N d and V range (n v / N v -0.5), so that the distance and speed ranges of the two-dimensional MUSIC spectrum estimation are [0, D range ] and [-V range , V range ] / 2, where: The speed sampling point is set to take into account the existence of positive and negative speeds in practical applications. However, the setting of the distance axis sampling point makes the perceived spectrum peak usually appear on one side of the two-dimensional MUSIC spectrum, which will affect the training and prediction effects of the model. Let β = d max / D range ≤1, the distance axis sampling point setting by formula (10) can make the two-dimensional MUSIC spectrum P music The position of the spectrum peak on the distance axis is [1-β, 1+β]N d / 2, that is, the probability of appearing in the center of the two-dimensional spectrum increases;

[0045] Step 2.16: Obtain the two-dimensional MUSIC spectrum P music Input the trained model to obtain the normalized perceptual parameter set and

[0046] Step 2.17: Convert the obtained perception parameter set into the distance and speed corresponding to the perception target. The conversion method is shown in the following formula (11):

[0047]

[0048] According to the above formula (11), the estimation result of the perception parameter is obtained and

[0049] Step 3: Model training process

[0050] Step 3.1: Set parameter D range , V range , d max , v max , the number of perceived targets is set to Set the distance ambiguity η d and velocity ambiguity η v , indicating the similarity of the perception parameters of the two targets, generates the distances d1 and v1 corresponding to the first perception target, satisfying 0≤d1≤d max and -v max≤v1≤v max , the distance and speed corresponding to the second perceived target are calculated according to the ambiguity as follows (12):

[0051]

[0052] Step 3.2: Generate training data according to step C1, generate 20,000 samples of size N d ×N v =40×40 2D-MUSIC spectrum is used as training data to train the model. The center positions of the two peaks in the training data are randomly distributed on the MUSIC spectrum, and the velocity ambiguity η v Uniformly distributed between 0 and 0.5, the distance ambiguity η d Uniformly distributed between 0 and 0.9, 5000 2D-MUSIC spectrum data generated in the same way were generated as a validation set. In the test set, the target range ambiguity increased from 0 to 0.9 in steps of 0.1. Other parameters were the same as the training set. 5000 data were generated for each range ambiguity, that is, the total size of the test set was 50000.

[0053] Step 3.3: Use the mean squared error function as the loss function during model training:

[0054]

[0055] In formula (13), represents the normalized perceptual parameter set, and has the following formula (14):

[0056]

[0057] In formula (14), k = 1, 2, and the parameters are optimized using the Adam optimizer; the batch size is set to 128; the weight decay rate is set to 0.00001; the initial learning rate is set to 0.001, and when the validation set error decreases by no more than one percent within three rounds, the learning rate is divided by 10; the model is trained for a total of 60 rounds.

[0058] The superior technical effects of the method of the present invention are as follows:

[0059] 1. The method described in the present invention utilizes the good image processing ability of the residual network and the feature selection ability of the attention mechanism to process the two-dimensional MUSIC perception spectrum image, achieving higher accuracy in separating similar targets.

[0060] 2. The method described in the present invention models the estimation problem of perception parameters as a regression prediction problem and directly outputs the perception parameters. Compared with the multi-classification indirect estimation method of dividing the grid and predicting the spectral peak position, the perception performance of this method is not limited by the grid resolution.

[0061] 3. The method of the present invention adjusts the sampling mode of the two-dimensional MUSIC spectrum so that the spectrum peaks are symmetrically distributed near the center of the two-dimensional spectrum, thereby accelerating the model convergence and ultimately enhancing the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] FIG1 is a schematic diagram of a system structure involved in the method of the present invention;

[0063] Figure 2 3 is a schematic diagram comparing the root mean square error of distance estimation of the method described in the present invention and other super-resolution perception methods under different test set signal-to-noise ratios.

[0064] FIG3 is a schematic diagram showing a comparison of the root mean square error of distance / speed estimation of the present invention and other super-resolution perception methods under different distance ambiguities. DETAILED DESCRIPTION

[0065] In order to more clearly understand the principles and features of the present invention, the specific implementation of the method of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0066] Example

[0067] As shown in FIG. 1 , the method, as shown in FIG. 1 , includes:

[0068] Step A, network model construction:

[0069] Step A1, the first layer of the neural network consists of a common convolution layer and a pooling layer. The convolution kernel size of the convolution layer is 3×3, the stride is 1, and the number of output channels is 64. A batch normalization layer is added after the convolution layer, and the ReLu function is used as the output function of the batch normalization layer. The maximum pooling layer is placed after the ReLu layer, with a kernel size of 3×3 and a stride of 2.

[0070] Step A2: The second and third neural networks are each composed of a classic residual block. Each residual block contains two convolutional layers. The convolution kernel size of each convolutional layer is set to 3×3, the step size is 1, the output channel of the second residual block is 64, and the output channel of the third residual block is 128.

[0071] In step A3, the fourth and fifth neural networks add an attention mechanism based on the residual block. The internal parameter settings of the fourth and fifth residual blocks are the same as those of the first two layers. The number of output channels is set to 256 and 512 respectively. A layer of SENet (Squeeze-and-excitation net) is added after the residual block, and the dimensionality reduction ratios are set to 8 and 16 respectively.

[0072] Step A4: After the fifth layer, a global average pooling layer is used to reduce the dimension of the data, followed by a fully connected layer as the output layer. The number of output neurons is set to in is the number of perceived targets;

[0073] Step B, signal processing flow:

[0074] Step B1, set parameters N, M and Where: N is the number of OFDM symbols, M is the number of subcarriers, is the modulation order, and the source generates a length of The serial bit data stream is processed The digital modulation of the order is mapped to MN constellation symbols, where: m = 0, ..., M-1, n = 0, ..., N-1;

[0075] Step B2: First, the MN constellation symbols are placed on the time-frequency domain grid by serial-to-parallel conversion to obtain the symbols to be sent in the time-frequency domain, which are recorded as A[m,n]. The transmitter performs M-point discrete inverse Fourier transform (Invers e Discret e Fourier Transform (IDFT) is used to obtain discrete time domain samples, which are then converted into continuous time domain signals to be transmitted x(t) through the transmit pulse shaping function, as shown in the following equation (1):

[0076]

[0077] In formula 1, Δf is the subcarrier spacing, T s is the duration of an OFDM symbol with a cyclic prefix (CP), satisfying T s =T+T cp , and T = 1 / Δf is the duration of the OFDM symbol information part, T cp is the duration of the CP part, g tx (t) is the transmission pulse shaping function;

[0078] Step B3: Setting perception parameters ρ k , d k and v k ,in is the number of perceived targets, ρ k , d k and v k are the channel gain, distance and speed corresponding to the kth target respectively, and there is

[0079] Step B4: constructing a time-delay domain channel impulse response based on the perception parameters where δ(·) is the Dirac delta function, τ k and v k are the time delay and Doppler corresponding to the kth path, satisfying τ k =2d k / c and v k =2v k f c / c, c represents the speed of light, f c Indicates the center carrier frequency of the transmitted signal;

[0080] Step B5: The transmitted signal reaches the sensing receiving end through the sensing channel, and the received signal r(t) is obtained by integrating the transmitted signal x(t) and the time-delay domain channel impulse response h(t, τ), as shown in the following formula (2):

[0081]

[0082] In formula (2), n(t) is the additive white Gaussian noise received by the receiver. The receiving end can obtain a discrete received signal by sampling the received signal at a time interval of T / M, removing the part of the sampled data that belongs to the cyclic prefix, and performing DFT on every M sampling points, that is, obtaining the received signal matrix Y[m, n] in the time-frequency domain: as shown in formula (3)

[0083]

[0084] The time-frequency domain receiving matrix is ​​regarded as the result of multiplying the transmitted symbol A[m,n] by the channel complex gain plus the noise, then Y=A⊙H(d,v)+N, where ⊙ represents the Hadamard product, H(d,v) is the channel impulse response matrix in the time-frequency domain, and Represents the perceptual parameter vector to be estimated, let represents the channel gain vector, then H(d, v) is further decomposed into H(d, v) = U(d)diag(Γ)B(v) T , where U(d) and B(v) represent the phase shift matrices corresponding to the delay and Doppler, respectively, as shown in equation (4):

[0085]

[0086] In formula (4)

[0087] Step B6: The estimation algorithm needs to receive channel-related information in the signal matrix. The received matrix Y needs to be divided by the transmitted symbol A by element-wise division to obtain the matrix O to be estimated:

[0088]

[0089] In the above formula (5), the row direction of the estimated matrix O contains the distance information of each target, and the column direction contains the speed information of each target. represents the noise matrix;

[0090] Step B7, rearrange O, splice the column vectors together one by one, put the distance information and speed information on the same dimension, and define the turning vector in: represents the Kronecker product, then the rearranged O is expressed as follows:

[0091]

[0092] In formula (6), represents the rearranged vector, and vec(·) represents the column vectorization operation;

[0093] Step B8, transpose q with its conjugate H Multiply and calculate the covariance matrix Q xx =qq H ;

[0094] Step B9, calculate the covariance matrix Q xx The eigenvalues ​​and corresponding eigenvectors of Q xx The eigenvalue decomposition Q xx =EΛE H , where: Λ=diag(λ1,...,λ MN ), E = [e1, ..., e MN ],λ l and e l They are the covariance matrices Q xx The eigenvalues ​​and corresponding eigenvectors of

[0095] Step B10, sort the calculated eigenvalues ​​in descending order so that the sorted eigenvalue sequence satisfies in The eigenvectors are also rearranged in the order of eigenvalues, and the rearranged eigenvector matrix is ​​recorded as

[0096] Step B11, the largest front The eigenvalues ​​are considered The eigenvalues ​​corresponding to the reflected signals of the targets, and the rest The eigenvalues ​​are considered to be noise eigenvalues. According to the covariance matrix Q xxThe received signal is decomposed into signal subspace and noise subspace according to the eigenvalue decomposition and sorting results. The result of subspace decomposition is expressed as follows:

[0097]

[0098] In formula (7), and Represent the signal subspace and the noise subspace respectively. Based on the above decomposition operation, the two-dimensional MUSIC spectrum is expressed as follows (8):

[0099]

[0100] In formula (8), d and v represent distance and speed respectively, c(d, v) is the steering vector defined in step 2.7, represents the noise subspace, (·) H represents the conjugate transpose.

[0101] Step B12, setting parameter N d and N v , respectively represent the number of sampling points when sampling the two-dimensional MUSIC spectrum along the distance axis and the velocity axis;

[0102] Step B13, based on the turning vectors u(d) and b(v) described in step B5, calculate the maximum distance range D estimated by the two-dimensional MUSIC algorithm range and maximum speed range V range , respectively satisfying and That is, we have the following formula (9):

[0103]

[0104] Step B14: confirm the maximum distance d that the perceived target can reach in the algorithm application scenario max and the maximum rate v max , satisfying 0≤d k ≤d max and -v max ≤v k ≤v max ,in: At the same time, it is necessary to ensure that the waveform parameters are set so that D range ≥d max and V range ≥v max / 2;

[0105] Step B15: based on the parameters set or calculated in the above three steps, the two-dimensional MUSIC spectrum is sampled along the distance axis and the speed axis. The sampled two-dimensional discrete MUSIC spectrum is denoted as P music, as shown in formula (10):

[0106]

[0107] In formula (10), n d =0, ..., N d , n v =0, ..., N v The usual practice is to set each sampling point to D range n d / N d and V range (n v / N v -0.5), so that the distance and speed ranges of the two-dimensional MUSIC spectrum estimation are [0, D range ] and [-V range , V range ] / 2, where: The speed sampling point is set to take into account the existence of positive and negative speeds in practical applications. However, the setting of the distance axis sampling point makes the perceived spectrum peak usually appear on one side of the two-dimensional MUSIC spectrum, which will affect the training and prediction effects of the model. Let β = d max / D range ≤1, the distance axis sampling point setting in formula (10) can make the two-dimensional MUSIC spectrum P music In the distance axis, the peak of the spectrum appears at [1-β, 1+β]N d / 2, that is, the probability of appearing in the center of the two-dimensional spectrum increases;

[0108] Step B16, the obtained two-dimensional MUSIC spectrum P music Input the trained model to obtain the normalized perceptual parameter set and

[0109] Step B17, converting the obtained perception parameter set into the distance and speed corresponding to the perception target, and the conversion method is shown in the following formula (11):

[0110]

[0111] According to the above formula (11), the estimation result of the perception parameter is obtained and

[0112] Step C, model training process:

[0113] Step C1, set parameter D range , V range , d max , v max, the number of perceived targets is set to Set the distance ambiguity η d and velocity ambiguity η v , indicating the similarity of the perception parameters of the two targets, generates the distances d1 and v1 corresponding to the first perception target, satisfying 0≤d1≤d max and -v max ≤v1≤v max , the distance and speed corresponding to the second perceived target are calculated according to the ambiguity as follows (12):

[0114]

[0115] Step C2: Generate training data according to step C1, generate 20,000 training data of size N. d ×N v =40×40 2D-MUSIC spectrum is used as training data to train the model. The center positions of the two peaks in the training data are randomly distributed on the MUSIC spectrum, and the velocity ambiguity η v Uniformly distributed between 0 and 0.5, the distance ambiguity η d Uniformly distributed between 0 and 0.9, 5000 2D-MUSIC spectrum data generated in the same way were generated as a validation set. In the test set, the target range ambiguity increased from 0 to 0.9 in steps of 0.1. Other parameters were the same as the training set. 5000 data were generated for each range ambiguity, that is, the total size of the test set was 50000.

[0116] Step C3, use the mean square error function as the loss function during model training:

[0117]

[0118] In formula (13), represents the normalized perceptual parameter set, and has the following formula (14):

[0119]

[0120] In formula (14), k = 1, 2, and the parameters are optimized using the Adam optimizer; the batch size is set to 128; the weight decay rate is set to 0.00001; the initial learning rate is set to 0.001, and when the validation set error decreases by no more than one percent within three rounds, the learning rate is divided by 10; the model is trained for a total of 60 rounds.

[0121] Verification Example

[0122] In order to verify the effect of the present invention, the following simulation is carried out to verify the performance of the present invention by taking a dual-target perception scenario as an example. The simulation verification takes a dual-target perception scenario as an example, setting the carrier frequency to 27.5 GHz, the subcarrier spacing to 120 kHz, the number of subcarriers to 16, and the number of OFDM symbols to 8. The maximum distance of the perceived target is d max = 600m, maximum speed is v max =100m / s.

[0123] Depend on Figure 2 It can be seen that with the increase of the signal-to-noise ratio of the test set, the estimated root mean square error (RMSE) of each model gradually decreases. Among them, when the signal-to-noise ratio of the training data is 0dB, the estimated RMSE of the method is first greater than the deep neural network method and the artificial neural network method, and then exceeds the deep neural network method and the artificial neural network method as the signal-to-noise ratio increases. This is because the method regards the estimation problem as a regression problem, and the estimation accuracy is not affected by the grid point resolution. Therefore, it shows higher performance under high signal-to-noise ratio, while it is greatly affected by noise under low signal-to-noise ratio. When the training data is 20dB, the method achieves the best effect when the signal-to-noise ratio exceeds 7dB due to its excellent feature extraction ability. In order to study the impact of similar targets on model performance, the distance difference and speed difference of the two targets in Figure 3 are set to be less than the resolution, and simulation is carried out. Among them, the distance ambiguity of the two targets in the training set is evenly distributed between 0 and 0.9, the speed ambiguity is evenly distributed between 0 and 0.5, and the signal-to-noise ratio of the training set is set to 20dB. By Figure 2 As shown in Figure 3, as the distance ambiguity increases, the estimation error of the traditional MUSIC algorithm first increases and then decreases. This is because when the ambiguity is small, the spectral peaks corresponding to the two targets are separated on the 2D-MUSIC spectrum. The increase in ambiguity increases the interference between the spectral peaks and increases the estimation error. When the ambiguity is greater than a certain threshold, the distance between the two spectral peaks is less than the resolution of the MUSIC algorithm, and the spectral peaks merge, leaving only one synthetic spectral peak. At this time, the estimation result is affected by this unique synthetic spectral peak, and the RMSE gradually decreases as the distance from the real spectral peak decreases. Compared with deep neural network and artificial neural network methods, the method of regression prediction of perception parameters in this method shows higher estimation accuracy in the simulation results. While maintaining an RMSE similar to that of the residual network model, this method reduces the number of model layers and the time spent on training, while also providing faster inference speed.

[0124] The present invention is not limited by the above embodiments. The above embodiments and descriptions are only for explaining the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, which are within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A synaesthesia integrated waveform super-resolution perception method based on residual network, comprising: Step 1, network model construction; Step 2, signal processing flow; Step 3: Model training process.

2. According to the residual network-based synaesthesia integrated waveform super-resolution perception method of claim 1, the network model construction in step 1 includes the following steps: Step 1.1 The first layer of the neural network consists of a common convolutional layer and a pooling layer. The convolutional layer has a kernel size of 3×3, a stride of 1, and 64 output channels. A batch normalization layer is added after the convolutional layer, and the ReLu function is used as the output function of the batch normalization layer. The maximum pooling layer is placed after the ReLu layer, with a kernel size of 3×3 and a stride of 2. Step 1.2: The second and third layers of the neural network are each composed of a classic residual block. Each residual block contains two convolutional layers. The convolution kernel size of each convolutional layer is set to 3×3, the step size is 1, the output channel of the second residual block is 64, and the output channel of the third residual block is 128; Step 1.3: The fourth and fifth neural network layers add an attention mechanism based on the residual block. The internal parameter settings of the fourth and fifth residual blocks are the same as the first two layers. The number of output channels is set to 256 and 512 respectively. A layer of SENet is added after the residual block, and the dimensionality reduction ratio is set to 8 and 16 respectively. Step 1.4: After the fifth layer, a global average pooling layer is used to reduce the dimension of the data, followed by a fully connected layer as the output layer. The number of output neurons is set to in is the number of perceived targets.

3. According to the residual network-based synaesthesia integrated waveform super-resolution perception method of claim 1, the signal processing flow in step 2 comprises the following steps: Step 2.1: Set the parameters N, M and Where: N is the number of OFDM symbols, M is the number of subcarriers, is the modulation order, and the source generates a length of The serial bit data stream is processed The digital modulation of the order is mapped to MN constellation symbols, where: m = 0, ..., M-1, n = 0, ..., N-1; Step 2.2: The MN constellation symbols are first placed on the time-frequency domain grid through serial-to-parallel transformation to obtain the symbols to be transmitted in the time-frequency domain, which are recorded as A[m, n]. The transmitter performs M-point discrete inverse Fourier transform (IDFT) on each row of A[m, n] to obtain discrete time domain samples, and then converts them into continuous time domain signals to be transmitted x(t) through the transmission pulse shaping function, as shown in the following formula (1): In formula 1, Δf is the subcarrier spacing, T s is the duration of an OFDM symbol with a cyclic prefix (CP), satisfying T s =T+T cp , and T = 1 / Δf is the duration of the OFDM symbol information part, T cp is the duration of the CP part, g tx (t) is the transmission pulse shaping function; Step 2.3: Set perception parameters ρ k , d k and v k ,in is the number of perceived targets, ρ k , d k and v k are the channel gain, distance and speed corresponding to the kth target respectively, and there is Step 2.4: Construct the time-delay domain channel impulse response based on the perception parameters where δ(·) is the Dirac delta function, τ k and v k are the time delay and Doppler corresponding to the kth path, satisfying τ k =2d k / c and v k =2v k f c / c, c represents the speed of light, f c Indicates the center carrier frequency of the transmitted signal; Step 2.5: The transmitted signal reaches the sensing receiver through the sensing channel. The received signal r(t) is obtained by integrating the transmitted signal x(t) and the time-delay domain channel impulse response h(t, τ), as shown in the following equation (2): In formula (2), n(t) is the additive white Gaussian noise received by the receiver. The receiving end can obtain a discrete received signal by sampling the received signal at a time interval of T / M, removing the part of the sampled data that belongs to the cyclic prefix, and performing DFT on every M sampling points, that is, obtaining the received signal matrix Y[m, n] in the time-frequency domain: as shown in formula (3) The time-frequency domain receiving matrix is ​​regarded as the result of multiplying the transmitted symbol A[m,n] by the channel complex gain plus the noise, then Y=A⊙H(d,v)+N, where ⊙ represents the Hadamard product, H(d,v) is the channel impulse response matrix in the time-frequency domain, and Represents the perceptual parameter vector to be estimated, let represents the channel gain vector, then H(d, v) is further decomposed into H(d, v) = U(d)diag(Γ)B(v) T , where U(d) and B(v) represent the phase shift matrices corresponding to the delay and Doppler, respectively, as shown in equation (4): In formula (4), Step 2.6: The estimation algorithm needs to receive channel-related information in the signal matrix. The received matrix Y needs to be divided by the transmitted symbol A by element-wise division to obtain the estimated matrix O: In the above formula (5), the row direction of the estimated matrix O contains the distance information of each target, and the column direction contains the speed information of each target. represents the noise matrix; Step 2.7: Rearrange O so that the column vectors are joined together one by one, placing the distance information and speed information on the same dimension, and defining the turning vector in: represents the Kronecker product, then the rearranged O is expressed as follows: In formula (6), represents the rearranged vector, and vec(·) represents the column vectorization operation; Step 2.8: Transpose q with its conjugate H Multiply and calculate the covariance matrix Q xx =qq H ; Step 2.9: Calculate the covariance matrix Q xx The eigenvalues ​​and corresponding eigenvectors of Q xx The eigenvalue decomposition Q xx =EΛE H , where: Λ=diag(λ1,...,λ MN ), E = [e1, ..., e MN ],λ l and e l They are the covariance matrices Q xx The eigenvalues ​​and corresponding eigenvectors of Step 2.10: Sort the calculated eigenvalues ​​in descending order so that the sorted eigenvalue sequence satisfies in The eigenvectors are also rearranged in the order of eigenvalues, and the rearranged eigenvector matrix is ​​recorded as Step 2.11: Maximum front The eigenvalues ​​are considered The eigenvalues ​​corresponding to the reflected signals of the targets, and the rest The eigenvalues ​​are considered to be noise eigenvalues. According to the covariance matrix Q xx The received signal is decomposed into signal subspace and noise subspace according to the eigenvalue decomposition and sorting results. The result of subspace decomposition is expressed as follows: In formula (7), and Represent the signal subspace and the noise subspace respectively. Based on the above decomposition operation, the two-dimensional MUSIC spectrum is expressed as follows (8): In formula (8), d and v represent distance and speed respectively, c(d, v) is the steering vector defined in step 2.7, represents the noise subspace, (·) H represents the conjugate transpose. Step 2.12: Set parameter N d and N v , respectively represent the number of sampling points when sampling the two-dimensional MUSIC spectrum along the distance axis and the velocity axis; Step 2.13: Based on the steering vectors u(d) and b(v) described in step B5, calculate the maximum distance range D estimated by the two-dimensional MUSIC algorithm range and maximum speed range V range , respectively satisfying and That is, we have the following formula (9): Step 2.14: Determine the maximum distance d that the target can reach in the algorithm application scenario max and the maximum rate v max , satisfying 0≤d k ≤d max and -v max ≤v k ≤v max ,in: At the same time, it is necessary to ensure that the waveform parameters are set so that D range ≥d max and V range ≥v max / 2; Step 2.15: According to the parameters set or calculated in the above three steps, the two-dimensional MUSIC spectrum is sampled along the distance axis and the velocity axis. The sampled two-dimensional discrete MUSIC spectrum is recorded as P music , as shown in formula (10): In formula (10), n d =0, ..., N d , n v =0, ..., N v The usual practice is to set each sampling point to D range n d / N d and V range (n v / N v -0.5), so that the distance and speed ranges of the two-dimensional MUSIC spectrum estimation are [0, D range ] and [-V range , V range ] / 2, where: The speed sampling point is set to take into account the existence of positive and negative speeds in practical applications. However, the setting of the distance axis sampling point makes the perceived spectrum peak usually appear on one side of the two-dimensional MUSIC spectrum, which will affect the training and prediction effects of the model. Let β = d max / D range ≤1, the distance axis sampling point setting by formula (10) can make the two-dimensional MUSIC spectrum P music The position of the spectrum peak on the distance axis is [1-β, 1+β]N d / 2, that is, the probability of appearing in the center of the two-dimensional spectrum increases; Step 2.16: Obtain the two-dimensional MUSIC spectrum P music Input the trained model to obtain the normalized perceptual parameter set and Step 2.17: Convert the obtained perception parameter set into the distance and speed corresponding to the perception target. The conversion method is shown in the following formula (11): According to the above formula (11), the estimation result of the perception parameter is obtained and 4. According to the residual network-based synaesthesia integrated waveform super-resolution perception method of claim 1, the model training process in step 3 comprises the following steps: Step 3.1: Set parameter D range , V range , d max , v max , the number of perceived targets is set to Set the distance ambiguity η d and velocity ambiguity η v , indicating the similarity of the perception parameters of the two targets, generates the distances d1 and v1 corresponding to the first perception target, satisfying 0≤d1≤d max and -v max ≤v1≤v max , the distance and speed corresponding to the second perceived target are calculated according to the ambiguity as follows (12): Step 3.2: Generate training data according to step C1, generate 20,000 samples of size N d ×N v =40×40 2D-MUSIC spectrum is used as training data to train the model. The center positions of the two peaks in the training data are randomly distributed on the MUSIC spectrum, and the velocity ambiguity η v Uniformly distributed between 0 and 0.5, the distance ambiguity η d Uniformly distributed between 0 and 0.9, 5000 2D-MUSIC spectrum data generated in the same way were generated as a validation set. In the test set, the target range ambiguity increased from 0 to 0.9 in steps of 0.

1. Other parameters were the same as the training set. 5000 data were generated for each range ambiguity, that is, the total size of the test set was 50000. Step 3.3: Use the mean squared error function as the loss function during model training: In formula (13), represents the normalized perceptual parameter set, and has the following formula (14): In formula (14), k = 1, 2, and the parameters are optimized using the Adam optimizer; the batch size is set to 128; the weight decay rate is set to 0.00001; the initial learning rate is set to 0.001, and when the validation set error decreases by no more than one percent within three rounds, the learning rate is divided by 10; the model is trained for a total of 60 rounds.

Images