A neural network design method for generating wireless signals for human perception

Abstract

The present invention discloses a neural network design method for generating human-perceived wireless signals, belonging to the field of deep learning for wireless signal processing. The method comprises the following steps: S1: converting the wireless signal input into a neural network into a spectrogram format to obtain a signal; S2: extracting local frequency features from the wireless signal using a one-dimensional Fourier transform network; S3: further extracting time-domain features from the wireless signal using a ViT network; S4: performing two-dimensional frequency-domain feature extraction and extracting global frequency features from the wireless signal using a two-dimensional Fourier transform network; and S5: inputting the extracted time-frequency features of the wireless signal into a signal generation network for the final output signal. The present invention provides a neural network design method for generating human-perceived wireless signals, which enables the generated wireless signal to have richer and more realistic frequency information, thereby improving the quality of the generated signal.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology for wireless signal processing, and in particular to a neural network design method for generating human-perceived wireless signals. Background Art

[0002] Wireless human sensing technology uses wireless signals (such as Wi-Fi, millimeter waves, Bluetooth, and ultra-wideband) to detect human activity, behavior, and physiological status. By processing wireless signals and analyzing changes in their propagation characteristics near the human body, it extracts information related to human behavior. It has widespread applications in smart homes, health monitoring, security surveillance, and other fields.

[0003] With the rapid development of artificial intelligence (AI) technology in recent years, AI-based wireless signal processing methods have gained widespread application. However, training AI models requires a large amount of wireless signal data, and acquiring high-quality wireless signals is extremely expensive. Therefore, generating simulated wireless signals has become a promising solution. This method not only significantly reduces signal acquisition costs but also provides rich training data for the model, thereby improving the efficiency and accuracy of wireless signal processing.

[0004] Thanks to the powerful capabilities of generative AI in generating content, a large body of work has been conducted using this technology to generate wireless signals. Existing wireless signal generation methods primarily utilize convolutional neural networks (CNNs) to extract features from original signals and then generate new simulated wireless signals based on these features. However, CNNs have limitations in extracting and generating wireless signal features, primarily due to their inadequate processing of the signal's time-series and frequency characteristics. Wireless signals are inherently time-series data and contain rich dynamic information, but CNNs struggle to capture these long-term dependencies, resulting in inadequate reproduction of time-domain features. Furthermore, frequency features in wireless signals, such as Doppler shift, are crucial for detecting target velocity, but the local convolution operations of CNNs struggle to extract the signal's global spectral information, resulting in frequency-domain distortion in the generated signal. These shortcomings result in suboptimal accuracy and fidelity in wireless signal generation using CNNs.

[0005] In summary, in order to generate high-quality wireless signals, there is an urgent need for a neural network that can effectively extract the time domain and frequency domain features of wireless signals at the same time, improve the accuracy and authenticity of signal generation, and thus meet the needs of high-precision applications. Summary of the Invention

[0006] The purpose of the present invention is to provide a neural network design method for generating human-perceived wireless signals. This method can effectively solve the problems in the prior art of not fully considering the rich physical characteristics of wireless signals, making it difficult to extract time-frequency characteristics from wireless signals, and making it difficult to generate high-precision and high-fidelity wireless signals.

[0007] To achieve the above objectives, the present invention provides a neural network design method for human perception of wireless signal generation, comprising the following steps:

[0008] S1. Convert the wireless signal input into the neural network into a spectrum format to obtain the signal , indicating that the input wireless signal format is a three-dimensional matrix vector with a length of M, a width of N, and a number of channels C;

[0009] S2. Perform one-dimensional frequency domain feature extraction and use a one-dimensional Fourier transform network to extract local frequency features in the wireless signal;

[0010] S3. Perform time domain feature extraction and use the ViT network to further extract the time domain features of the wireless signal; including:

[0011] S31, extracting basic time domain features;

[0012] S32, extracting enhanced time domain features;

[0013] S4, performing two-dimensional frequency domain feature extraction, using a two-dimensional Fourier transform network to extract global frequency features in the wireless signal;

[0014] S5. Input the extracted time-frequency features of the wireless signal into a signal generation network for final output signal.

[0015] Preferably, the content of S2 is as follows:

[0016] S21, one-dimensional Fourier transform and feature extraction;

[0017] S22. Perform adaptive feature extraction.

[0018] Preferably, the content of S21 is as follows:

[0019] S211. Perform discrete Fourier transform along the length latitude of the wireless signal. The specific formula is as follows:

[0020] ;

[0021] in Indicates signal No. List, represents the imaginary unit, express exist The spectrum at

[0022] S212, apply the operation in S211 to each channel to obtain a spectrum diagram .

[0023] Preferably, the content of S22 is as follows:

[0024] S221, Randomly initialize the learnable parameter vector , copy and extend it along the width direction and channel direction to obtain the parameter vector ;

[0025] S222, Randomly initialize the learnable parameter vector , copy and extend it along the length direction and channel direction to obtain the parameter vector ;

[0026] S223, the parameter vector in S221 and the parameter vector in S222 Multiply to get the final learnable parameter vector ;

[0027] S224, the learnable parameter vector in S223 Spectrum diagram with S22 Element-by-element product to obtain the filtered spectrum , the specific formula is as follows:

[0028] ;

[0029] S225, the spectrum after filtering in S224 Perform inverse discrete Fourier transform to obtain one-dimensional frequency domain features ;

[0030] in, Represents the filtered signal No. Column No. elements, Representing a spectrum graph No. Column No. elements;

[0031] S226, apply the operation in S225 to each channel to obtain a filtered signal .

[0032] Preferably, the calculation formula of S225 is as follows:

[0033] .

[0034] Preferably, S31, extracting basic time domain features is as follows:

[0035] S311, initialize the two-dimensional convolution module , setting its kernel size parameter to , setting its step size parameter to , set its channel number parameter to ;

[0036] S312, the two-dimensional convolution module in S311 Applied to S226 filter signal Get the feature vector , the specific formula is as follows:

[0037] ;

[0038] in, ;

[0039] Preferably, the content of the enhanced time domain feature extraction in S32 is as follows:

[0040] S321, initialize the two-dimensional convolution module , setting its kernel size parameter to , setting its step size parameter to , set its expansion 0 parameter to , set its channel number parameter to ;

[0041] S322, the two-dimensional convolution module in S321 Applied to S226 filter signal Get the feature vector , the specific formula is as follows:

[0042] ;

[0043] in, .

[0044] Preferably, the content of S4 is as follows:

[0045] S41, shape reshaping:

[0046] S411, the feature vector in S312 Reshape to get the reshaped feature vector ;

[0047] S412, the feature vector in S322 Reshape to get the reshaped feature vector ;

[0048] S42, layer normalization:

[0049] S421, reshape the feature vector in S411 Each row along the feature dimension Calculate the mean , with the first Behavior example, the specific formula is as follows:

[0050] ;

[0051] in, Indicates the The mean of the rows, Represents the feature vector No. The first elements;

[0052] S422, reshape the feature vector in S411 Each row along the feature dimension Calculating variance , with the first Behavior example, the specific formula is as follows:

[0053] ;

[0054] in, express No. Variance of rows;

[0055] S423, reshape the feature vector in S411 Each element of is normalized to obtain the normalized feature vector , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0056] ;

[0057] S424, reshape the feature vector in S412 Perform the same operations from S421 to S423 to obtain the normalized feature vector ;

[0058] S43, two-dimensional Fourier transform and feature extraction:

[0059] Normalized eigenvectors in S423 Perform two-dimensional Fourier transform processing to obtain the spectrum diagram ;

[0060] S44, Adaptive feature extraction:

[0061] S441, Randomly initialize the learnable parameter vector ;

[0062] S442, the learnable parameter vector in S441 Spectrum diagram with S43 Element-by-element product to obtain the filtered spectrum , the specific formula is as follows:

[0063] ;

[0064] S443, the spectrum after filtering in S442 Perform two-dimensional inverse discrete Fourier transform to obtain two-dimensional frequency domain features , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0065] ;

[0066] S444, reshape the feature vector in S424 Perform the same operations from S441 to S443 to obtain the two-dimensional frequency domain features ;

[0067] S45, shape reshaping:

[0068] S451, reshape the feature vector in S443 Reshape to get feature vector ;

[0069] S452, reshape the feature vector in S444 Reshape to get feature vector ;

[0070] S46, layer normalization:

[0071] S461, the characteristic vector in S451 Each row along the feature dimension Calculate the mean , with the first Behavior example, the specific formula is as follows:

[0072] ;

[0073] in, Indicates the The mean of the rows, Represents the feature vector No. The first elements;

[0074] S462, the characteristic vector in S451 Each row along the feature dimension Calculating variance , with the first Behavior example, the specific formula is as follows:

[0075] ;

[0076] in, express No. Variance of rows;

[0077] S463, the characteristic vector in S451 Each element of is normalized to obtain the eigenvector , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0078] ;

[0079] S464, the characteristic vector in S452 Perform the same operations from S461 to S463 to obtain the feature vector .

[0080] Preferably, the calculation formula of S43 is as follows:

[0081] First Rank Take column elements as an example,

[0082] ;

[0083] Among them, the spectrum .

[0084] Preferably, the content of S5 is as follows:

[0085] S51. Unify input dimensions:

[0086] S511, using the neural network fully connected layer The eigenvector in S463 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0087] ;

[0088] in, ;

[0089] S512, using the neural network fully connected layer The eigenvector in S464 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0090] ;

[0091] in, ;

[0092] S52, Feature Recovery:

[0093] S521, unified feature vector in S511 Perform all the same operations as S4 , and the restored eigenvector is obtained , the specific formula is as follows:

[0094] ;

[0095] S522, unify the feature vector in S512 Perform all the same operations as S4 , and the restored eigenvector is obtained , the specific formula is as follows:

[0096] ;

[0097] S53, output dimension unification:

[0098] S531. Using the fully connected layer of the neural network The restored feature vector in S521 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0099] ;

[0100] in, ;

[0101] S532, using the neural network fully connected layer The restored feature vector in S522 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0102] ;

[0103] in, ;

[0104] S54, output generation signal:

[0105] S541, unify the feature vector in S531 Reshape to get feature vector ;

[0106] S542, unify the feature vector in S532 Reshape to get feature vector ;

[0107] S543, the feature vector in S541 and the eigenvector in S542 The final wireless signal is generated by adding bit by bit along the center point .

[0108] Therefore, the present invention adopts the above-mentioned neural network design method for human perception of wireless signal generation, which has the following beneficial effects:

[0109] (1) It can extract effective frequency domain features for wireless signal generation from the original wireless signal. Compared with traditional signal generation methods, by integrating one-dimensional Fourier transform and two-dimensional Fourier transform into the neural network, and enabling the neural network to learn and extract effective frequency features in an adaptive way, the generated wireless signal has richer and more realistic frequency information.

[0110] (2) It can extract effective time domain features for wireless signal generation from the original wireless signal. Based on the ViT architecture, a global receptive field is established through the self-attention mechanism, enabling the neural network to capture long-distance temporal relationships. At the same time, it can adaptively learn and extract effective time domain features, thereby improving the quality of the generated signal.

[0111] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 This is a flowchart of wireless signal feature extraction and generation proposed in an embodiment of the present invention;

[0113] Figure 2 A schematic diagram of inputting wireless signals according to an embodiment of the present invention;

[0114] Figure 3 Schematic diagram of the wireless signal generated by the wireless signal generation method proposed in the embodiment of the present invention and the real wireless signal; wherein (a) is the real wireless signal , (b) is to generate wireless signal . DETAILED DESCRIPTION

[0115] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0116] See also Figure 1 , a neural network design method for generating human-sensing wireless signals, comprising the following steps:

[0117] S1, such as Figure 2 , convert the wireless signal input into the neural network into a spectrum format and obtain the signal , indicating that the input wireless signal format is a three-dimensional matrix vector with a length of M, a width of N, and a number of channels C.

[0118] S2. Perform one-dimensional frequency domain feature extraction and use a one-dimensional Fourier transform network to extract local frequency features in the wireless signal.

[0119] S21. One-dimensional Fourier transform and feature extraction. The content is as follows:

[0120] S211. Perform discrete Fourier transform along the length latitude of the wireless signal. The specific formula is as follows:

[0121] ;

[0122] in Indicates signal No. List, represents the imaginary unit, express exist The spectrum at .

[0123] S212, apply the operation in S211 to each channel to obtain a spectrum diagram .

[0124] S22. Perform adaptive feature extraction.

[0125] S221, Randomly initialize the learnable parameter vector , copy and extend it along the width direction and channel direction to obtain the parameter vector ;

[0126] S222, Randomly initialize the learnable parameter vector , copy and extend it along the length direction and channel direction to obtain the parameter vector ;

[0127] S223, the parameter vector in S221 and the parameter vector in S222 Multiply to get the final learnable parameter vector ;

[0128] S224, the learnable parameter vector in S223 Spectrum diagram with S22 Element-by-element product to obtain the filtered spectrum , the specific formula is as follows:

[0129] ;

[0130] S225, the spectrum after filtering in S224 Perform inverse discrete Fourier transform to obtain one-dimensional frequency domain features , the specific formula is as follows:

[0131] ;

[0132] in, Represents the filtered signal No. Column No. elements, Representing a spectrum graph No. Column No. elements;

[0133] S226, apply the operation in S225 to each channel to obtain a filtered signal .

[0134] S3. Perform time domain feature extraction and use the ViT network to further extract the time domain features of the wireless signal.

[0135] S31. Basic time domain feature extraction:

[0136] S311, initialize the two-dimensional convolution module , setting its kernel size parameter to , setting its step size parameter to , set its channel number parameter to ;

[0137] S312, the two-dimensional convolution module in S311 Applied to S226 filter signal Get the feature vector , the specific formula is as follows:

[0138] ;

[0139] in, ;

[0140] S32, Enhanced time domain feature extraction:

[0141] S321, initialize the two-dimensional convolution module , setting its kernel size parameter to , setting its step size parameter to , set its expansion 0 parameter to , set its channel number parameter to ;

[0142] S322, the two-dimensional convolution module in S321 Applied to S226 filter signal Get the feature vector , the specific formula is as follows:

[0143] ;

[0144] in, .

[0145] S4. Perform two-dimensional frequency domain feature extraction and use a two-dimensional Fourier transform network to extract global frequency features in the wireless signal.

[0146] S41, shape reshaping:

[0147] S411, the feature vector in S312 Reshape to get the reshaped feature vector ;

[0148] S412, the feature vector in S322 Reshape to get the reshaped feature vector ;

[0149] S42, layer normalization:

[0150] S421, reshape the feature vector in S411 Each row along the feature dimension Calculate the mean , with the first Behavior example, the specific formula is as follows:

[0151] ;

[0152] in, Indicates the The mean of the rows, Represents the feature vector No. The first elements;

[0153] S422, reshape the feature vector in S411 Each row along the feature dimension Calculating variance , with the first Behavior example, the specific formula is as follows:

[0154] ;

[0155] in, express No. Variance of rows;

[0156] S423, reshape the feature vector in S411 Each element of is normalized to obtain the normalized feature vector , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0157] ;

[0158] S424, reshape the feature vector in S412 Perform the same operations from S421 to S423 to obtain the normalized feature vector ;

[0159] S43, two-dimensional Fourier transform and feature extraction:

[0160] Normalized eigenvectors in S423 Perform two-dimensional Fourier transform processing to obtain the spectrum diagram , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0161] ;

[0162] Among them, the spectrum ;

[0163] S44, Adaptive feature extraction:

[0164] S441, Randomly initialize the learnable parameter vector ;

[0165] S442, the learnable parameter vector in S441 Spectrum diagram with S43 Element-by-element product to obtain the filtered spectrum , the specific formula is as follows:

[0166] ;

[0167] S443, the spectrum after filtering in S442 Perform two-dimensional inverse discrete Fourier transform to obtain two-dimensional frequency domain features , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0168] ;

[0169] S444, reshape the feature vector in S424 Perform the same operations from S441 to S443 to obtain the two-dimensional frequency domain features ;

[0170] S45, shape reshaping:

[0171] S451, reshape the feature vector in S443 Reshape to get feature vector ;

[0172] S452, reshape the feature vector in S444 Reshape to get feature vector ;

[0173] S46, layer normalization:

[0174] S461, the characteristic vector in S451 Each row along the feature dimension Calculate the mean , with the first Behavior example, the specific formula is as follows:

[0175] ;

[0176] in, Indicates the The mean of the rows, Represents the feature vector No. The first elements;

[0177] S462, the characteristic vector in S451 Each row along the feature dimension Calculating variance , with the first Behavior example, the specific formula is as follows:

[0178] ;

[0179] in, express No. Variance of rows;

[0180] S463, the characteristic vector in S451 Each element of is normalized to obtain the eigenvector , with the first Rank Taking column elements as an example, the specific formula is as follows:

[0181] ;

[0182] S464, the characteristic vector in S452 Perform the same operations from S461 to S463 to obtain the feature vector .

[0183] S5. Input the extracted time-frequency features of the wireless signal into a signal generation network for final output signal.

[0184] S51. Unify input dimensions:

[0185] S511, using the neural network fully connected layer The eigenvector in S463 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0186] ;

[0187] in, ;

[0188] S512, using the neural network fully connected layer The eigenvector in S464 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0189] ;

[0190] in, ;

[0191] S52, Feature Recovery:

[0192] S521: Unify the eigenvectors in S511 Perform all the same operations as S4 , and the restored eigenvector is obtained , the specific formula is as follows:

[0193] ;

[0194] S522, unify the feature vector in S512 Perform all the same operations as S4 , and the restored eigenvector is obtained , the specific formula is as follows:

[0195] ;

[0196] S53, output dimension unification:

[0197] S531. Using the fully connected layer of the neural network The restored feature vector in S521 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0198] ;

[0199] in, ;

[0200] S532, using the neural network fully connected layer The restored feature vector in S522 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows:

[0201] ;

[0202] in, ;

[0203] S54, output generation signal:

[0204] S541, unify the feature vector in S531 Reshape to get feature vector ;

[0205] S542, unify the feature vector in S532 Reshape to get feature vector ;

[0206] S543, the feature vector in S541 and the eigenvector in S542 The final wireless signal is generated by adding bit by bit along the center point ,like Figure 3 shown.

[0207] The test was conducted on two public wireless signal datasets, the Dop-NET dataset and the IURHA2023 dataset. The experiment was run three times with different random seeds, and the final average and standard deviation results were reported. The specific results are as follows:

[0208] like Figure 2 As shown, the wireless signal generated by S533 on the IURHA2023 dataset On real signal Visually, they are extremely similar;

[0209] As shown in Table 1, the wireless signal generated by this embodiment shows improvements in average peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indicators compared with the traditional method.

[0210] Table 1 Comparison of signal quality generated by the neural network of this embodiment and traditional neural network

[0211] ;

[0212] Therefore, the present invention adopts the above-mentioned neural network design method for human perception of wireless signal generation. By embedding a learnable frequency domain feature extraction module in the neural network, features are adaptively extracted from the wireless signal for generating refined wireless signals. Its frequency domain feature extraction module includes one-dimensional frequency domain feature extraction and two-dimensional frequency domain feature extraction. The characteristic of adaptive extraction is that the neural network can change and gradually optimize the extracted features by continuously iteratively updating the neural network parameters, thereby improving the quality of the generated wireless signal.

[0213] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A neural network design method for human perception of wireless signal generation, characterized in that: The human perception of wireless signals is to realize the perception of human activities, behaviors and physiological states, by processing wireless signals, analyzing the changes in the propagation characteristics of the wireless signals near the human body, and extracting information related to human behavior, including the following steps: S1. Convert the wireless signal input into the neural network into a spectrum format to obtain the signal , indicating that the input wireless signal format is a three-dimensional matrix vector with a length of M, a width of N, and a number of channels C; S2. Perform one-dimensional frequency domain feature extraction and use a one-dimensional Fourier transform network to extract local frequency features in the wireless signal; The content is as follows: S21, one-dimensional Fourier transform and feature extraction; S22, performing adaptive feature extraction; S22 content is as follows: S221, Randomly initialize the learnable parameter vector , copy and extend it along the width direction and channel direction to obtain the parameter vector ; S222, Randomly initialize the learnable parameter vector , copy and extend it along the length direction and channel direction to obtain the parameter vector ; S223, the parameter vector in S221 and the parameter vector in S222 Multiply to get the final learnable parameter vector ; S224, the learnable parameter vector in S223 Spectrum diagram with S22 Element-by-element product to obtain the filtered spectrum , the specific formula is as follows: ； S225, the spectrum after filtering in S224 Perform inverse discrete Fourier transform to obtain one-dimensional frequency domain features The calculation formula is as follows: ； in, Represents the filtered signal No. Column No. elements, Representing a spectrum graph No. Column No. elements; S226, apply the operation in S225 to each channel to obtain a filtered signal ; S3. Perform time domain feature extraction and use the Vision Transformer network to further extract the time domain features of the wireless signal; including: S31. Extract basic time domain features; the content is as follows: S311, initialize the two-dimensional convolution module , setting its kernel size parameter to , setting its step size parameter to , set its channel number parameter to ; S312, the two-dimensional convolution module in S311 Applied to S226 filter signal Get the feature vector , the specific formula is as follows: ； in, , ; S32, extracting enhanced time domain features; the content is as follows: S321, initialize the two-dimensional convolution module , setting its kernel size parameter to , setting its step size parameter to , set its expansion 0 parameter to , set its channel number parameter to ; S322, the two-dimensional convolution module in S321 Applied to S226 filter signal Get the feature vector , the specific formula is as follows: ； in, , ; S4, performing two-dimensional frequency domain feature extraction, using a two-dimensional Fourier transform network to extract global frequency features in the wireless signal; S5. Input the extracted time-frequency features of the wireless signal into a signal generation network for final output signal.

2. A neural network design method for generating human-perceived wireless signals according to claim 1, characterized in that: S21 content is as follows: S211. Perform discrete Fourier transform along the length latitude of the wireless signal. The specific formula is as follows: ； in Indicates signal No. List, represents the imaginary unit, express exist The spectrum at S212, apply the operation in S211 to each channel to obtain a spectrum diagram .

3. The neural network design method for generating human-perceived wireless signals according to claim 2, wherein: S4 content is as follows: S41, shape reshaping: S411, the feature vector in S312 Reshape to get the reshaped feature vector ; S412, the feature vector in S322 Reshape to get the reshaped feature vector ; S42, layer normalization: S421, reshape the feature vector in S411 Each row along the feature dimension Calculate the mean , with the first Behavior example, the specific formula is as follows: ； in, Indicates the The mean of the rows, Represents the feature vector No. The first elements; S422, reshape the feature vector in S411 Each row along the feature dimension Calculating variance , with the first Behavior example, the specific formula is as follows: ； in, express No. Variance of rows; S423, reshape the feature vector in S411 Each element of is normalized to obtain the normalized feature vector , with the first Rank Taking column elements as an example, the specific formula is as follows: ； S424, reshape the feature vector in S412 Perform the same operations from S421 to S423 to obtain the normalized feature vector ; S43, two-dimensional Fourier transform and feature extraction: Normalized eigenvectors in S423 Perform two-dimensional Fourier transform processing to obtain the spectrum diagram ; Among them, the spectrum ; S44, Adaptive feature extraction: S441, Randomly initialize the learnable parameter vector ; S442, the learnable parameter vector in S441 Spectrum diagram with S43 Element-by-element product to obtain the filtered spectrum , the specific formula is as follows: ； S443, the spectrum after filtering in S442 Perform two-dimensional inverse discrete Fourier transform to obtain two-dimensional frequency domain features , with the first Rank Taking column elements as an example, the specific formula is as follows: ； S444, reshape the feature vector in S424 Perform the same operations from S441 to S443 to obtain the two-dimensional frequency domain features ; S45, shape reshaping: S451, reshape the feature vector in S443 Reshape to get feature vector ; S452, reshape the feature vector in S444 Reshape to get feature vector ; S46, layer normalization: S461, the characteristic vector in S451 Each row along the feature dimension Calculate the mean , with the first Behavior example, the specific formula is as follows: ； in, Indicates the The mean of the rows, Represents the feature vector No. The first elements; S462, the characteristic vector in S451 Each row along the feature dimension Calculating variance , with the first Behavior example, the specific formula is as follows: ； in, express No. Variance of rows; S463, the characteristic vector in S451 Each element of is normalized to obtain the eigenvector , with the first Rank Taking column elements as an example, the specific formula is as follows: ； S464, the characteristic vector in S452 Perform the same operations from S461 to S463 to obtain the feature vector .

4. A neural network design method for generating human-perceived wireless signals according to claim 3, characterized in that: The S43 calculation formula is as follows: First Rank Take column elements as an example, ； Among them, the spectrum .

5. The neural network design method for generating human-perceived wireless signals according to claim 4, wherein: S5 content is as follows: S51. Unify input dimensions: S511, using the neural network fully connected layer The eigenvector in S463 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows: ； in, ; S512, using the neural network fully connected layer The eigenvector in S464 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows: ； in, ; S52, Feature Recovery: S521: Unify the eigenvectors in S511 Perform all the same operations as S4 , and the restored eigenvector is obtained , the specific formula is as follows: ； S522, unify the feature vector in S512 Perform all the same operations as S4 , and the restored eigenvector is obtained , the specific formula is as follows: ； S53, output dimension unification: S531. Using the fully connected layer of the neural network The restored feature vector in S521 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows: ； in, ; S532, using the neural network fully connected layer The restored feature vector in S522 Map to a unified dimension to obtain a unified feature vector , the specific formula is as follows: ； in, ; S54, output generation signal: S541, unify the feature vector in S531 Reshape to get feature vector ; S542, unify the feature vector in S532 Reshape to get feature vector ; S543, the feature vector in S541 and the eigenvector in S542 The final wireless signal is generated by adding bit by bit along the center point .

Images