Corrected deep learning underwater channel estimation method, electronic equipment and medium

Abstract

The invention discloses a modified deep learning underwater channel estimation method, electronic equipment and a medium. The method comprises the following steps: acquiring original signal data, and acquiring initial response of a channel based on initial estimation of least square; aiming at the initial response, carrying out noise reduction processing through a transform domain filtering algorithm to obtain noise reduction data; inputting the noise reduction data into a deep learning network, learning nonlinear characteristics of a channel, and obtaining a channel estimation value; and correcting the channel estimation value according to the time-frequency attention network to obtain a channel correction value. According to the method, key time-frequency information is highlighted through adaptive weighted fusion of the time-frequency information, unimportant time-frequency information is suppressed, and fine correction of a channel estimation result is realized.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and more specifically, to a modified deep learning underwater channel estimation method, electronic equipment, and medium. Background Art

[0002] Due to the complex and variable time-frequency characteristics of underwater OFDM signals, traditional deep learning models struggle to adaptively focus on key time-frequency regions, limiting estimation performance in complex multipath and strong noise environments. Furthermore, if the algorithm neglects to effectively suppress noise during the initial estimation phase, it will negatively impact the subsequent training and estimation of the deep learning model.

[0003] Therefore, it is necessary to develop a modified deep learning underwater channel estimation method, electronic equipment and medium.

[0004] The information disclosed in the background technology section of the present invention is only intended to deepen the understanding of the general background technology of the present invention, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0005] The present invention proposes a modified deep learning underwater channel estimation method, electronic device and medium, which can highlight key time-frequency information and suppress unimportant time-frequency information through adaptive weighted fusion of time-frequency information, thereby achieving refined correction of channel estimation results.

[0006] In a first aspect, an embodiment of the present disclosure provides a modified deep learning underwater channel estimation method, comprising:

[0007] Obtain the original signal data and obtain the preliminary response of the channel based on the initial estimation of least squares;

[0008] For the preliminary response, noise reduction is performed using a transform domain filtering algorithm to obtain noise-reduced data;

[0009] The denoised data is input into the deep learning network to learn the nonlinear characteristics of the channel and obtain the channel estimation value;

[0010] The channel estimation value is corrected according to the time-frequency attention network to obtain the channel correction value.

[0011] Preferably, performing noise reduction processing on the preliminary response by using a transform domain filtering algorithm, and obtaining noise reduction data includes:

[0012] The preliminary response is transformed into the transform domain through DFT, and G in the transform domain M (q) is the value corresponding to the preliminary response in the frequency domain;

[0013] Set the cutoff frequency p cLow-pass filter, obtaining a low-pass filter sequence;

[0014] Perform IDFT on the low-pass filtered sequence to obtain the denoised data after variation domain denoising.

[0015] Preferably, the transform domain sequence expression is:

[0016]

[0017] Preferably, the low-pass filtering sequence is:

[0018]

[0019] Among them, p c It is 1 / 2 of the cyclic prefix length.

[0020] Preferably, the noise reduction data is:

[0021]

[0022] Preferably, correcting the channel estimation value according to the time-frequency attention network to obtain the channel correction value includes:

[0023] Obtain frequency-domain attention-weighted data through the frequency-domain attention branch;

[0024] Obtain the time-weighted attention data through the time-weighted attention branch;

[0025] Add the frequency-domain attention-weighted data and the time-domain attention-weighted data to obtain fused data;

[0026] The fused data is linearly transformed and converted back into a one-dimensional shape as the output, which is the channel correction value.

[0027] Preferably, obtaining frequency-domain attention-weighted data through the frequency-domain attention branch includes:

[0028] The Conv2D layer is used for frequency domain feature extraction. This layer has 32 filters and a convolution kernel size of (3,1), that is, 3 subcarriers × 1 time point. Sliding convolution is performed in the frequency domain dimension. The activation function is relu and the padding method is the same:

[0029] relu(x)=max(0,x)

[0030] The input feature map is For the kth convolution kernel and bias b_k, the output at frequency domain position i is calculated as:

[0031]

[0032] Among them, F is the number of frequency domain points, T is the time step, i is the frequency domain position, m is the offset of the convolution kernel in the frequency domain, c is the channel index, and the input data is x. The convolution operation formula is:

[0033]

[0034] Where w is the convolution kernel weight, b is the bias, and f is the activation function;

[0035] The frequency domain attention weights are generated through another Conv2D layer. This layer has 2 filters, the convolution kernel size is (3, 1), the activation function is sigmoid, and the input is z. The sigmoid function formula is:

[0036]

[0037] Since the output range of the sigmoid function is between 0 and 1, the generated weights represent the importance of different frequency domain features:

[0038]

[0039] The channel estimation value is multiplied by the frequency domain attention weight through the Multiply operation to obtain the frequency domain attention weighted data.

[0040] Preferably, obtaining the time-domain attention-weighted data through the time-domain attention branch includes:

[0041] The data dimension is converted to (1, 52, 2) through the Permute operation, and the original frequency domain dimension is used as the width and the time domain dimension as the height, so as to achieve sliding in the time step direction and perform sliding convolution operation in the time domain;

[0042] The Conv2D layer is used for time domain feature extraction. The layer has 32 filters, the convolution kernel size is (1,3), that is, 1 subcarrier × 3 time points, the activation function is relu, the padding method is the same, and the input after dimension conversion is Convolution kernel The output at time step t is:

[0043]

[0044] Where t is the time step position, n is the frequency domain offset (0, 1, 2), c is the channel index, T is the time step, and F is the number of frequency domain points;

[0045] Use the Permute operation to convert the data dimension back to (52, 1, 2) and generate the time domain attention weight through the Conv2D layer;

[0046] The channel estimation value is multiplied by the time domain attention weight through the Multiply operation to obtain the data weighted by the time domain attention.

[0047] In a second aspect, an embodiment of the present disclosure further provides an electronic device, the electronic device comprising:

[0048] a memory storing executable instructions;

[0049] A processor runs the executable instructions in the memory to implement the modified deep learning underwater channel estimation method.

[0050] In a third aspect, an embodiment of the present disclosure further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, implements the modified deep learning underwater channel estimation method.

[0051] The methods and apparatus of the present invention have other features and advantages that will be apparent from, or will be described in detail in, the accompanying drawings and subsequent detailed descriptions incorporated herein, which together serve to explain the specific principles of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0053] Figure 1 A flowchart showing the steps of modifying a deep learning underwater channel estimation method according to an embodiment of the present invention is shown.

[0054] Figure 2 A schematic diagram showing the principle of a transform domain denoising algorithm according to an embodiment of the present invention is shown.

[0055] Figure 3 A schematic diagram showing the comparison of signal-to-noise ratios before and after transform domain filtering according to an embodiment of the present invention is shown.

[0056] Figure 4 A schematic diagram of a TFAN model network architecture according to an embodiment of the present invention is shown.

[0057] Figure 5 A schematic diagram of a modified DNN data processing flow according to an embodiment of the present invention is shown.

[0058] Figure 6A schematic diagram showing a comparison of normalized variances of various algorithms for a simulation channel according to an embodiment of the present invention is shown.

[0059] Figure 7 A schematic diagram showing a comparison of bit error rates of various algorithms for a simulation channel according to an embodiment of the present invention is shown.

[0060] Figure 8 A schematic diagram showing a comparison of constellation diagrams of various algorithms for a simulation channel according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0061] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0062] Figure 1 A flowchart showing the steps of modifying a deep learning underwater channel estimation method according to an embodiment of the present invention is shown.

[0063] like Figure 1 As shown, the modified deep learning underwater channel estimation method includes:

[0064] Step 101: Obtain original signal data and obtain a preliminary response of the channel based on an initial estimation using least squares.

[0065] Step 102: performing noise reduction processing on the preliminary response using a transform domain filtering algorithm to obtain noise-reduced data;

[0066] Step 103: input the denoised data into a deep learning network to learn the nonlinear characteristics of the channel and obtain a channel estimation value;

[0067] Step 104: Correct the channel estimation value according to the time-frequency attention network to obtain a channel correction value.

[0068] In one example, noise reduction processing is performed on the preliminary response using a transform domain filtering algorithm, and obtaining noise-reduced data includes:

[0069] The preliminary response is transformed into the transform domain through DFT, and G in the transform domain M (q) is the value corresponding to the preliminary response in the frequency domain;

[0070] Set the cutoff frequency p c Low-pass filter, obtaining a low-pass filter sequence;

[0071] Perform IDFT on the low-pass filtered sequence to obtain the denoised data after variation domain denoising.

[0072] In one example, the transform domain sequence expression is:

[0073]

[0074] In one example, the low-pass filtering sequence is:

[0075]

[0076] Among them, p c It is 1 / 2 of the cyclic prefix length.

[0077] In one example, the denoised data is:

[0078]

[0079] In one example, correcting the channel estimation value based on the time-frequency attention network to obtain the channel correction value includes:

[0080] Obtain frequency-domain attention-weighted data through the frequency-domain attention branch;

[0081] Obtain the time-weighted attention data through the time-weighted attention branch;

[0082] Add the frequency-domain attention-weighted data and the time-domain attention-weighted data to obtain fused data;

[0083] The fused data is linearly transformed and converted back into a one-dimensional shape as the output, which is the channel correction value.

[0084] In one example, obtaining frequency-domain attention-weighted data through the frequency-domain attention branch includes:

[0085] The Conv2D layer is used for frequency domain feature extraction. This layer has 32 filters and a convolution kernel size of (3,1), that is, 3 subcarriers × 1 time point. Sliding convolution is performed in the frequency domain dimension. The activation function is relu and the padding method is the same:

[0086] relu(x)=max(0,x)

[0087] The input feature map is For the kth convolution kernel and bias b_k, the output at frequency domain position i is calculated as:

[0088]

[0089] Among them, F is the number of frequency domain points, T is the time step, i is the frequency domain position, m is the offset of the convolution kernel in the frequency domain, c is the channel index, and the input data is x. The convolution operation formula is:

[0090]

[0091] Where w is the convolution kernel weight, b is the bias, and f is the activation function;

[0092] The frequency domain attention weights are generated through another Conv2D layer. This layer has 2 filters, the convolution kernel size is (3, 1), the activation function is sigmoid, and the input is z. The sigmoid function formula is:

[0093]

[0094] Since the output range of the sigmoid function is between 0 and 1, the generated weights represent the importance of different frequency domain features:

[0095]

[0096] The channel estimation value is multiplied by the frequency domain attention weight through the Multiply operation to obtain the frequency domain attention weighted data.

[0097] In one example, obtaining temporal attention-weighted data through the temporal attention branch includes:

[0098] The data dimension is converted to (1, 52, 2) through the Permute operation, and the original frequency domain dimension is used as the width and the time domain dimension as the height, so as to achieve sliding in the time step direction and perform sliding convolution operation in the time domain;

[0099] The Conv2D layer is used for time domain feature extraction. The layer has 32 filters, the convolution kernel size is (1,3), that is, 1 subcarrier × 3 time points, the activation function is relu, the padding method is the same, and the input after dimension conversion is Convolution kernel The output at time step t is:

[0100]

[0101] Where t is the time step position, n is the frequency domain offset (0, 1, 2), c is the channel index, T is the time step, and F is the number of frequency domain points;

[0102] Use the Permute operation to convert the data dimension back to (52, 1, 2) and generate the time domain attention weight through the Conv2D layer;

[0103] The channel estimation value is multiplied by the time domain attention weight through the Multiply operation to obtain the data weighted by the time domain attention.

[0104] Specifically, the received signal is the starting point of the entire algorithm, obtaining the original signal data from the receiver. This original signal contains the channel information that needs to be estimated, but is also mixed with interference factors such as noise.

[0105] Initial Estimation Based on Least Squares (LS): The least squares (LS) algorithm uses the received signal to estimate the initial channel response. This stage yields a preliminary channel response estimate that contains noise. Because the least squares algorithm does not specifically address noise, this estimate is subject to some noise interference.

[0106] Transform Domain Denoising Processing (TDDP): After obtaining a preliminary channel response estimate based on LS, a TDDP filtering algorithm is used for noise reduction. In the TDDP, signals and noise often have different characteristics. By manipulating coefficients in the TDDP, noise can be effectively suppressed while preserving the key signal characteristics, improving the data quality of the input to the deep learning network (DNN). Clean data helps the DNN better learn the characteristics of the channel and avoid being misled by noise.

[0107] The noise sources in underwater acoustic communication environments are complex and extensive. When these noises are superimposed on the signal, the estimation error of the traditional LS algorithm increases significantly. If the frequency domain channel response estimation value H obtained by the LS algorithm is directly LS (k) Deep learning increases the difficulty of convergence and the complexity of model training. In severe cases, this can cause large fluctuations in the model's gradient calculations, leading to unstable training and even exploding or vanishing gradients. This can also mislead the DNN into learning incorrect features, severely impacting DNN training and estimation accuracy. Therefore, performing noise reduction preprocessing before DNN processing is crucial.

[0108] Figure 2 A schematic diagram showing the principle of a transform domain denoising algorithm according to an embodiment of the present invention is shown.

[0109] In view of the characteristics of underwater multipath fading channel environment, the channel response value changes slowly, the noise interference term changes relatively quickly and the energy is relatively concentrated, the algorithm of the present invention adopts the transform domain filtering algorithm to perform noise reduction preprocessing on the rough channel estimate before DNN. The algorithm principle is as follows Figure 2 shown. Figure 2 Medium H M (m) is the initial frequency domain response of the channel after the least squares algorithm (LS), and H M(m) is transformed into the transform domain through DFT, and G in the transform domain M (q) is the frequency domain H M The transform domain sequence can be regarded as the "spectrum" sequence of the channel frequency domain response, which can reflect the speed of change of the channel response value in the frequency domain. The transform domain sequence expression is:

[0110]

[0111] In G M In the (q) sequence, the noise component exists in the entire frequency range, while the useful channel information is usually at a "lower frequency" position. Therefore, a suitable cutoff frequency p can be set. c The low-pass filter sets the "higher frequency" part corresponding to the noise to zero to eliminate noise interference as much as possible and retain the useful information part.

[0112]

[0113] The algorithm of the present invention selects p c It is 1 / 2 of the cyclic prefix length.

[0114] Afterwards, Perform IDFT to obtain the frequency domain channel response value after variation domain denoising

[0115]

[0116] Figure 3 A schematic diagram showing the comparison of signal-to-noise ratios before and after transform domain filtering according to an embodiment of the present invention is shown.

[0117] The signal-to-noise ratio comparison before and after transform domain denoising is as follows: Figure 3 As shown in the figure represents the transform domain before denoising, It represents the transform domain denoising. It can be seen that the signal-to-noise ratio of the channel response after transform domain denoising is significantly higher than that before denoising, and the data quality is better, which is more conducive to the DNN deep learning of its complex nonlinear characteristics of the channel and fitting the channel response value.

[0118] Nonlinear Feature Learning Based on DNN: Deep learning networks (DNNs) have powerful nonlinear fitting capabilities. At this stage, the denoised coarse channel estimate data is input into the DNN. Using a multi-layered neuron structure, the DNN extracts and learns features from the input data layer by layer. During training, the DNN continuously adjusts its internal weights and biases to learn the nonlinear characteristics of the channel. The training process typically uses a large amount of known channel data and the corresponding received signal data, optimizing the DNN parameters through a backpropagation algorithm. The DNN performs nonlinear fitting on the denoised coarse channel estimate data in the transform domain to obtain a fitted channel estimate. This estimate more accurately reflects the nonlinear characteristics of the channel than the previous preliminary estimate based on LS.

[0119] Estimation Result Correction Based on Time-Frequency Attention Network (TFAN): Time-Frequency Attention Network (TFAN) is a network structure that can focus on key time-frequency information. In the present invention, TFAN is used to fine-tune the results after DNN fitting. TFAN can adaptively assign weights according to the characteristics of the signal in the time-frequency domain, and pay more attention to those time-frequency information that are critical to channel estimation. The results after DNN fitting are processed by TFAN, focusing on key time-frequency information, and further improving the accuracy of channel estimation. The final result is a more accurate final channel estimation result (Final Channel Estimation Result) that has been finely corrected.

[0120] Figure 4 A schematic diagram of a TFAN model network architecture according to an embodiment of the present invention is shown.

[0121] Compared with LS and MMSE algorithms, DNN algorithm has greatly improved underwater channel estimation, but it is difficult for DNN deep learning model to adaptively focus on key time-frequency areas, resulting in limited estimation performance in complex multipath and strong noise environments. This paper proposes a modified DNN algorithm based on convolutional neural network time-frequency attention network (TFAN) corrector, which further refines the more accurate channel estimation results after DNN algorithm processing to further improve the channel estimation accuracy. The core idea of ​​the algorithm is to combine the attention mechanism in the time-frequency domain to efficiently extract key features from the input data, thereby improving the performance of channel estimation. The network architecture of the TFAN model is as follows: Figure 4 shown.

[0122] The present invention uses 32 filters in the Conv2D layer for feature extraction, which is a relatively suitable balance between model complexity and computational cost. When extracting features in the frequency domain, the present invention uses a convolution kernel size of (3,1), and when extracting features in the time domain, a convolution kernel size of (1,3). When performing weight generation and linear transformation, since the frequency domain response of the underwater acoustic channel is complex (including amplitude and phase), the Conv2D layer used for weight generation and linear transformation only needs to use 2 filters to process the real and imaginary parts of the complex signal respectively. The details are as follows:

[0123] (1) Frequency domain attention branch: suppressing severely faded subcarriers and enhancing the characteristics of stable subcarriers

[0124] a) Feature extraction: Frequency domain feature extraction is performed using the Conv2D layer. This layer has 32 filters and a convolution kernel size of (3,1), which is 3 subcarriers × 1 time point. Sliding convolution is performed in the frequency domain. The activation function is ReLU, and the padding method is the same.

[0125] relu(x) = max(0,x) (4)

[0126] Assume that the input feature map is (F is the number of frequency domain points, which is 52 in the present invention; T is the time step, which is 1 in the present invention), for the kth convolution kernel and bias b_k, the output at frequency domain position i is calculated as:

[0127]

[0128] Where i is the frequency domain position (0≤i<52), m is the offset of the convolution kernel in the frequency domain (0, 1, 2), c is the channel index (0 and 1, corresponding to 2 input channels), and the relu activation function introduces nonlinearity to retain positive features.

[0129] Assuming the input data is x, the convolution operation formula can be simplified to:

[0130]

[0131] Where w is the convolution kernel weight, b is the bias, and f is the activation function (here is the relu function). Through this layer of convolution operation, the network can extract frequency-domain related features from the input data.

[0132] b) Weight Generation: Next, frequency-domain attention weights are generated through another Conv2D layer. This layer has two filters, a convolution kernel size of (3, 1), and a sigmoid activation function. Let the input be z, and the sigmoid function formula is:

[0133]

[0134] Since the output range of the sigmoid function is between 0 and 1, the generated weights can represent the importance of different frequency domain features. The formula is expressed as:

[0135]

[0136] c) Weighted operation: Finally, the original data is multiplied by the frequency domain attention weight through the Multiply operation to obtain the data weighted by the frequency domain attention. Let the original data be x and the weight be w, then the weighted data is:

[0137] x weighted =x×w (9)

[0138] In this way, the network can reweight the data according to the importance of frequency domain features and highlight important frequency domain features.

[0139] (2) Temporal attention branch: suppressing the time steps of noisy bursts and enhancing the features of reliable time steps

[0140] a) Dimension conversion: First, the data dimension is converted to (1, 52, 2) through the Permute operation, with the original frequency domain dimension as the width and the time domain dimension as the height, so as to achieve sliding in the time step direction and perform sliding convolution operation in the time domain.

[0141] b) Feature extraction: Use the Conv2D layer for time domain feature extraction. This layer has 32 filters, the convolution kernel size is (1,3), that is, 1 subcarrier × 3 time points, the activation function is relu, and the padding method is the same. The convolution operation formula is similar to the frequency domain feature extraction. Assume that the input after dimension conversion is (T is the time step, F is the number of frequency domain points), convolution kernel The output at time step t is:

[0142]

[0143] Where t is the time step position, n is the frequency domain offset (0, 1, 2), and c is the channel index. Through this layer of convolution operation, the network can extract time-domain related features from the data.

[0144] c) Dimension recovery and weight generation: Use the Permute operation again to convert the data dimension back to (52, 1, 2), and then generate the time domain attention weight through the Conv2D layer. The weight generation method is the same as the frequency domain, using the sigmoid function.

[0145] d) Weighting operation: The original data is multiplied by the temporal attention weight through the Multiply operation to obtain the temporal attention-weighted data. In this way, the network can reweight the data according to the importance of temporal features, highlighting important temporal features.

[0146] (3) Feature fusion:

[0147] Add the data weighted by frequency domain and time domain attention, and the formula is expressed as:

[0148] x final =x freq +x time (11)

[0149] This can fuse important features in the frequency domain and time domain, enabling the network to comprehensively consider key information in the time and frequency domains.

[0150] (4) Linear transformation and output:

[0151] Finally, a linear transformation is performed through a Conv2D layer with two filters, a kernel size of (3, 1), a linear activation function (i.e., y = x), and the same padding. Finally, a Reshape operation is used to convert the data back to a one-dimensional shape of (104,) as the output, which is the final estimate of the channel frequency response.

[0152] Figure 5 A schematic diagram of a modified DNN data processing flow according to an embodiment of the present invention is shown.

[0153] The channel frequency domain response obtained by denoising preprocessing and then DNN processing is Assume that the final channel frequency domain response value after TFAN correction is H final (k), the data processing flow can be simplified as follows Figure 5 shown.

[0154] The present invention also provides an electronic device, which includes: a memory storing executable instructions; and a processor that runs the executable instructions in the memory to implement the above-mentioned modified deep learning underwater channel estimation method.

[0155] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned modified deep learning underwater channel estimation method.

[0156] To facilitate understanding of the solutions and effects of the embodiments of the present invention, three specific application examples are given below. Those skilled in the art should understand that these examples are only for facilitating understanding of the present invention, and any specific details thereof are not intended to limit the present invention in any way.

[0157] Example 1

[0158] The performance of the proposed channel estimation algorithm is simulated using a classic underwater acoustic channel, the Bellhop channel. The channel parameters are shown in Table 1.

[0159] Table 1 Channel parameters

[0160]

[0161]

[0162] Figure 6 A schematic diagram showing a comparison of normalized variances of various algorithms for a simulation channel according to an embodiment of the present invention is shown.

[0163] Figure 7 A schematic diagram showing a comparison of bit error rates of various algorithms for a simulation channel according to an embodiment of the present invention is shown.

[0164] Figure 8 A schematic diagram showing a comparison of constellation diagrams of various algorithms for a simulation channel according to an embodiment of the present invention is shown.

[0165] The performance of LS algorithm, MMSE algorithm, LS algorithm + transform domain denoising preprocessing + DNN algorithm (LS+NRP+DNN) and LS algorithm + transform domain denoising + DNN algorithm + TFAN algorithm (LS+NRP+DNN+TFAN) were simulated. The simulation results are as follows. Figure 6-Figure 8 shown.

[0166] Figure 6 middle, represents the LS algorithm, represents the MMSE algorithm, It represents LS algorithm + transform domain denoising preprocessing + DNN algorithm. It represents LS algorithm + transform domain denoising + DNN algorithm + TFAN algorithm. Figure 6The normalized mean square error (NMSE) of different algorithms under the underwater multipath fading simulation channel conditions modeled by MATLAB in the present invention shows that with the increase of signal-to-noise ratio (SNR), the NMSE of all algorithms shows a downward trend, and the performance gap between different algorithms also gradually increases. Among them, LS+NRP+DNN, that is, the NMSE performance of the preliminary improved algorithm of the present invention, is greatly improved compared with the LS and NMSE algorithms, and the improvement effect is more obvious with the increase of signal-to-noise ratio (SNR). LS+NRP+DNN+TFAN, that is, the NMSE performance of the final improved algorithm of the present invention, has the best performance among all the above algorithms. The final improved algorithm of the present invention further improves the accuracy of channel estimation on the basis of the preliminary improved algorithm of the present invention, and shows superior performance in the entire SNR range.

[0167] Figure 7 middle, means no channel estimation, represents the LS algorithm, represents the MMSE algorithm, It represents LS algorithm + transform domain denoising preprocessing + DNN algorithm. It represents LS algorithm + transform domain denoising + DNN algorithm + TFAN algorithm. Figure 7 The bit error rate (BER) of different algorithms under the conditions of underwater multipath fading simulation channels modeled using MATLAB by the present invention shows that LS+NRP+DNN, the preliminary improved algorithm of the present invention, shows significant performance improvement in BER performance compared to traditional LS and MMSE algorithms under different SNR conditions, especially in the medium and high SNR regions. The final improved algorithm of the present invention, LS+NRP+DNN+TFAN, further reduces the system's bit error rate (BER) based on the preliminary improved algorithm, improves the accuracy of channel estimation, and demonstrates superior performance across the entire SNR range.

[0168] right Figure 8 The constellation diagrams of the signals obtained by different algorithms under the underwater multipath fading simulation channel conditions modeled by MATLAB in the present invention are compared, and it can be found that LS+NRP+DNN+TFAN, that is, the final improved algorithm of the present invention ( Figure 8 The concentration of data points in the constellation diagram marked as Improved-Estimate is significantly higher than that of the LS and MMSE algorithms. The data points are more closely clustered in the ideal position, which is closest to the effect of the actual channel response value on the signal. This means that the final improved algorithm of the present invention can more effectively suppress noise and interference when processing signals, further improving the accuracy of the signal.

[0169] Simulation results show that the proposed channel estimation algorithm outperforms other traditional channel estimation algorithms in terms of normalized variance and bit error rate. This invention will provide a more efficient and accurate channel estimation solution for underwater OFDM communication systems, and has important theoretical and practical value in promoting the development of underwater communication technology.

[0170] Example 2

[0171] The present disclosure provides an electronic device, which includes: a memory storing executable instructions; and a processor running the executable instructions in the memory to implement the above-mentioned modified deep learning underwater channel estimation method.

[0172] An electronic device according to an embodiment of the present disclosure includes a memory and a processor.

[0173] The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), a hard disk, a flash memory, etc.

[0174] The processor may be a central processing unit (CPU) or other form of processing unit having data processing capability and / or instruction execution capability, and may control other components in the electronic device to perform desired functions. In one embodiment of the present disclosure, the processor is used to execute the computer-readable instructions stored in the memory.

[0175] Those skilled in the art should understand that in order to solve the technical problem of how to obtain a good user experience, this embodiment may also include well-known structures such as a communication bus and an interface, and these well-known structures should also be included in the scope of protection of this disclosure.

[0176] For detailed description of this embodiment, please refer to the corresponding description in the aforementioned embodiments, which will not be repeated here.

[0177] Example 3

[0178] An embodiment of the present disclosure provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the modified deep learning underwater channel estimation method.

[0179] According to an embodiment of the present disclosure, a computer-readable storage medium stores non-transitory computer-readable instructions, which, when executed by a processor, execute all or part of the steps of the aforementioned methods of the embodiments of the present disclosure.

[0180] The above-mentioned computer-readable storage media include, but are not limited to, optical storage media (e.g., CD-ROMs and DVDs), magneto-optical storage media (e.g., MOs), magnetic storage media (e.g., magnetic tapes or mobile hard disks), media with built-in rewritable non-volatile memory (e.g., memory cards), and media with built-in ROM (e.g., ROM cartridges).

[0181] Those skilled in the art should understand that the above description of the embodiments of the present invention is only for the purpose of illustrative purposes only to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any given examples.

[0182] While various embodiments of the present invention have been described above, the above description is intended to be illustrative, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A modified deep learning underwater channel estimation method, characterized in that: include: Obtain the original signal data and obtain the preliminary response of the channel based on the initial estimation of least squares; For the preliminary response, noise reduction is performed using a transform domain filtering algorithm to obtain noise-reduced data; The denoised data is input into the deep learning network to learn the nonlinear characteristics of the channel and obtain the channel estimation value; The channel estimation value is corrected according to the time-frequency attention network to obtain the channel correction value.

2. The modified deep learning underwater channel estimation method according to claim 1, wherein: For the preliminary response, noise reduction is performed using a transform domain filtering algorithm. The noise reduction data obtained includes: The preliminary response is transformed into the transform domain through DFT, and G in the transform domain M (q) is the value corresponding to the preliminary response in the frequency domain; Set the cutoff frequency p c Low-pass filter, obtaining a low-pass filter sequence; Perform IDFT on the low-pass filtered sequence to obtain the denoised data after variation domain denoising.

3. The modified deep learning underwater channel estimation method according to claim 2, wherein: The transform domain sequence expression is:

4. The modified deep learning underwater channel estimation method according to claim 2, wherein: The low-pass filter sequence is: Among them, p c It is 1 / 2 of the cyclic prefix length.

5. The modified deep learning underwater channel estimation method according to claim 2, wherein: The denoised data is:

6. The modified deep learning underwater channel estimation method according to claim 1, wherein: The channel estimation value is corrected according to the time-frequency attention network, and the channel correction value obtained includes: Obtain frequency-domain attention-weighted data through the frequency-domain attention branch; Obtain the time-weighted attention data through the time-weighted attention branch; Add the frequency-domain attention-weighted data and the time-domain attention-weighted data to obtain fused data; The fused data is linearly transformed and converted back into a one-dimensional shape as the output, which is the channel correction value.

7. The modified deep learning underwater channel estimation method according to claim 6, wherein: The frequency-domain attention-weighted data obtained through the frequency-domain attention branch include: The Conv2D layer is used for frequency domain feature extraction. This layer has 32 filters and a convolution kernel size of (3,1), that is, 3 subcarriers × 1 time point. Sliding convolution is performed in the frequency domain dimension. The activation function is relu and the padding method is the same: relu(x)=max(0,x) The input feature map is For the kth convolution kernel and bias b_k, the output at frequency domain position i is calculated as: Among them, F is the number of frequency domain points, T is the time step, i is the frequency domain position, m is the offset of the convolution kernel in the frequency domain, c is the channel index, and the input data is x. The convolution operation formula is: Where w is the convolution kernel weight, b is the bias, and f is the activation function; The frequency domain attention weights are generated through another Conv2D layer. This layer has 2 filters, the convolution kernel size is (3, 1), the activation function is sigmoid, and the input is z. The sigmoid function formula is: Since the output range of the sigmoid function is between 0 and 1, the generated weights represent the importance of different frequency domain features: The channel estimation value is multiplied by the frequency domain attention weight through the Multiply operation to obtain the frequency domain attention weighted data.

8. The modified deep learning underwater channel estimation method according to claim 6, wherein: The temporal attention-weighted data obtained through the temporal attention branch includes: The data dimension is converted to (1, 52, 2) through the Permute operation, and the original frequency domain dimension is used as the width and the time domain dimension as the height, so as to achieve sliding in the time step direction and perform sliding convolution operation in the time domain; The Conv2D layer is used for time domain feature extraction. The layer has 32 filters, the convolution kernel size is (1,3), that is, 1 subcarrier × 3 time points, the activation function is relu, the padding method is the same, and the input after dimension conversion is Convolution kernel The output at time step t is: Where t is the time step position, n is the frequency domain offset (0, 1, 2), c is the channel index, T is the time step, and F is the number of frequency domain points; Use the Permute operation to convert the data dimension back to (52, 1, 2) and generate the time domain attention weight through the Conv2D layer; The channel estimation value is multiplied by the time domain attention weight through the Multiply operation to obtain the data weighted by the time domain attention.

9. An electronic device, characterized in that: The electronic device comprises: a memory storing executable instructions; A processor that runs the executable instructions in the memory to implement the modified deep learning underwater channel estimation method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the modified deep learning underwater channel estimation method according to any one of claims 1 to 8.

Images