A spatial optical signal transmission method based on self-encoding compressed sensing

Abstract

The present invention discloses a spatial optical signal transmission method based on self-encoding compressed sensing. The method firstly modulates the amplitude of the transmitted radio frequency signal to a multi-wavelength optical carrier and mixes it with a pseudo-random sequence, and realizes random sub-sampling of the signal. Secondly, the randomly sub-sampled signal is coupled to the optical fiber through a collimator, and then the optical signal is converted into an electrical signal through a photoelectric converter to obtain a signal waveform image after compressive sampling. Then, an oscilloscope is used to collect the original radio frequency signal waveform image, and a signal data set is constructed to input the signal reconstruction network model constructed based on the autoencoder for pre-training. Finally, the waveform images after compressive sampling are combined into a data set, the pre-trained signal reconstruction network model is fine-tuned, and the reconstructed waveform image is output. The present invention effectively solves the problem of high computational cost and limited real-time performance of the reconstruction process, and can obtain higher quality reconstruction results under the same compression ratio.

Description

Technical Field

[0001] The present invention belongs to the technical field of spatial optical signal compressed sensing information reconstruction technology processing, and specifically relates to a spatial optical signal transmission method based on self-encoding compressed sensing. Background Art

[0002] Compressed sensing (CS), also known as compressed sampling, exploits the sparse characteristics of the signal and reconstructs the signal by randomly sampling the discrete samples of the signal at a sampling rate much lower than the Nyquist sampling rate. The theory of compressed sensing is widely used in information theory, image processing, earth science, optical / microwave imaging, pattern recognition, wireless communication, biomedical engineering and other fields. However, when the amount of data continues to increase, especially when it comes to spatial signal processing, traditional compressed sensing algorithms may face the problems of high computational complexity and insufficient real-time processing speed. Therefore, in recent years, the introduction of deep learning networks has provided new solutions to these problems. The CS reconstruction algorithm originates from computationally expensive convex / greedy algorithms, which cannot meet the real-time operation in demanding application fields. If deep learning and compressed sensing are combined, the deep learning model is used to learn the sparse representation of the signal, so that the original high-quality signal can be restored from a small amount of sampled data or sensor readings. Through the powerful representation learning ability of deep neural networks, more efficient signal acquisition, transmission and recovery can be achieved under the framework of compressed sensing, bringing new possibilities and performance improvements to fields such as spatial optical signal transmission. Contents of the invention

[0003] In order to solve the problem that the computational cost of compressed sensing reconstruction process is too high and cannot meet the problem of real-time data processing, the present invention provides a spatial optical signal transmission method based on autoencoder compressed sensing. The proposed method combines the compressed sensing reconstruction algorithm based on deep learning with the spatial optical compression sampling technology, which can restore the input signal well.

[0004] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0005] The technical solution includes two processes: compression sampling and reconstruction. The signal compression sampling process includes the following three steps:

[0006] S1. A transmitted RF signal is amplitude modulated to a multi-wavelength optical carrier by a Mach-Zehnder modulator.

[0007] S2. The modulated multi-wavelength optical carrier signal first passes through a diffraction grating to disperse and arrange the optical carrier signals of different wavelengths, and then passes through a spatial light modulator to achieve mixing of the signal with a pseudo-random sequence in the frequency domain. Random sub-sampling of the signal is achieved by controlling the amplitude screen loading rate of the spatial light modulator.

[0008] S3. The randomly sub-sampled signal is coupled into the optical fiber through a collimator, and then converted into an electrical signal through a photoelectric converter, which is low-pass filtered to obtain the signal waveform image after compressed sampling.

[0009] The signal is then reconstructed. The reconstruction process is divided into the following two steps:

[0010] S4. Use an oscilloscope to collect the original RF signal waveform image, construct a signal data set to input the signal reconstruction network model built based on the autoencoder, output the reconstructed waveform image, and perform pre-training.

[0011] S5. The waveform images compressed and sampled by S3 are combined into a data set, the pre-trained signal reconstruction network model is fine-tuned, the reconstructed waveform images are output, the spatial optical signal transmission is completed, and the performance is evaluated.

[0012] Furthermore, the process of mixing the modulated optical carrier signal with the pseudo-random sequence in the frequency domain is as follows:

[0013] The modulated multi-wavelength optical carrier signal first passes through a diffraction grating to disperse and arrange the optical carrier signals of different wavelengths, and then passes through a spatial light modulator to achieve mixing of the signal with a pseudo-random sequence in the frequency domain.

[0014] Furthermore, random sub-sampling of the signal is achieved by controlling the loading rate of the spatial light modulator amplitude screen, and the process is as follows: a two-dimensional pseudo-random signal with the same size as the spatial light modulator spot is loaded through the computer control center to control the amplitude of the corresponding pixel point on the amplitude screen, and the random sub-sampling of the signal is achieved by setting the speed of loading the amplitude screen on the computer. The two-dimensional pseudo-random signal is a binary distribution of "0" and "1", and corresponds to the "0" and "255" pixel amplitudes of the light modulation amplitude, respectively.

[0015] Furthermore, the implementation process of the signal reconstruction network model is as follows: the input waveform image passes through the initial reconstruction module and the hybrid reconstruction module in sequence, and the reconstructed waveform image is output.

[0016] The initial reconstruction module uses the initial reconstruction matrix to complete the initial reconstruction of the signal and performs automatic training through back propagation. Specifically, the randomly generated Gaussian matrix is ​​used as the initial reconstruction matrix and multiplied element by element with the input waveform image to obtain the initial reconstructed image. The hybrid reconstruction module is formed by integrating the Transformer backbone network based on the iterative shrinkage threshold algorithm ISTA and CNN.

[0017] The hybrid reconstruction module is composed of n cascaded restoration stages, and the input initial reconstructed image is passed through the hybrid reconstruction module to generate the final reconstructed photon signal; each stage of the hybrid reconstruction module is composed of a Transformer backbone network and convolutional layers added before and after the Transformer backbone network, and the Transformer backbone network structure includes an encoder and a decoder part, the encoder and the decoder are composed of fully connected layers, and a multi-layer self-attention mechanism is introduced after the encoder and the decoder to capture the local and global features of the image, and a soft threshold function is introduced before the encoder for processing, and reverse training is performed through the iterative shrinkage threshold algorithm ISTA.

[0018] The beneficial effects of the present invention are as follows:

[0019] The present invention has a simple structure and a novel reconstruction method. Compared with the existing compressed sensing spatial optical signal processing system, it can effectively solve the problems of high computational cost and limited real-time performance in the reconstruction process. At the same time, the reconstruction method uses ISTA and Transformer backbone network and CNN deep learning network algorithm signals for reconstruction, which can obtain higher quality reconstruction results under the same compression ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the sampling and reconstruction process of the present invention;

[0021] Figure 2 It is the experimental transmission effect diagram of the present invention. DETAILED DESCRIPTION

[0022] A typical embodiment of a spatial optical signal transmission method based on self-encoding compressed sensing of the present invention is described in detail below, and the method is further described in detail. It is necessary to point out here that the following embodiments are only used to further illustrate the method and cannot be understood as limiting the protection scope of the method. Personnel skilled in the art in the field can make some non-essential improvements and adjustments to the method based on the above content of the method, which still fall within the protection scope of the present invention.

[0023] The present invention proposes a spatial optical signal transmission method based on self-encoding compressed sensing, which includes two processes: signal compression sampling and reconstruction. The signal compression sampling process includes the following three steps:

[0024] S1. A transmitted RF signal is amplitude modulated to a multi-wavelength optical carrier by a Mach-Zehnder modulator.

[0025] S2. The modulated multi-wavelength optical carrier signal first passes through a diffraction grating to disperse and arrange the optical carrier signals of different wavelengths, and then passes through a spatial light modulator to achieve mixing of the signal with a pseudo-random sequence in the frequency domain. Random sub-sampling of the signal is achieved by controlling the amplitude screen loading rate of the spatial light modulator.

[0026] S3. The randomly sub-sampled signal is coupled into the optical fiber through a collimator, and then converted into an electrical signal through a photoelectric converter, which is low-pass filtered to obtain the signal waveform image after compressed sampling.

[0027] The signal is then reconstructed. The reconstruction process is divided into the following two steps:

[0028] S4. Use an oscilloscope to collect a large number of original RF signal waveform images, construct a signal data set, set the data set into training set, validation set, and test set in a 6:2:2 ratio, input the signal reconstruction network model built based on the autoencoder, output the reconstructed waveform image, perform pre-training, perform gradient descent on the training error, adjust the model hyperparameters, obtain a large number of sample learning networks, and conduct a preliminary evaluation of the network by comparing the reconstruction output results.

[0029] S5. The waveform images sampled by S3 are combined into a data set, the pre-trained signal reconstruction network model is fine-tuned, the reconstructed waveform images are output, the spatial optical signal transmission is completed, and the generalization ability of the model is evaluated through the reconstruction results to verify the effectiveness of the model.

[0030] The implementation process of the signal reconstruction network model is as follows: the input waveform image passes through the initial reconstruction module and the mixed reconstruction module in sequence.

[0031] The initial reconstruction module uses the initial reconstruction matrix to complete the initial reconstruction of the signal and performs automatic training through back propagation. Specifically, the randomly generated Gaussian matrix is ​​used as the initial reconstruction matrix and multiplied element by element with the input waveform image to obtain the initial reconstructed image.

[0032] The hybrid reconstruction module is formed by integrating a Transformer backbone network based on an iterative shrinkage threshold algorithm ISTA and a CNN.

[0033] The hybrid reconstruction module is composed of n cascaded restoration stages, and the input initial reconstructed image is passed through the hybrid reconstruction module to generate the final reconstructed photon signal; each stage of the hybrid reconstruction module is composed of a Transformer backbone network and convolutional layers added before and after the Transformer backbone network. The Transformer backbone network structure includes an encoder and a decoder part. The encoder and the decoder are composed of fully connected layers, and multi-layer self-attention mechanisms are introduced after the encoder and the decoder to capture the local and global features of the image, and a soft threshold function processing is introduced before the decoder.

[0034] Based on the deep learning compressed sensing reconstruction algorithm and the spatial light compression sampling method, the present invention proposes a compressed sensing spatial light signal transmission method based on an autoencoder, such as Figure 1 As shown, the system connects a single-mode optical fiber 2 through a light source 1 with different wavelengths, and the single-mode optical fiber 2 is serially connected to a Mach-Zehnder modulator 3, modulating the input signal into a corresponding wavelength optical signal, and finally transmitting it to the receiving end control center through an optical fiber 4. The control center is connected to a diffraction grating 5, and the grating 5 is serially connected to a spatial light modulator 6. The spatial light modulator 6 realizes the mixing of the signal in the frequency domain with the pseudo-random sequence. The spatial light modulator 6 is connected to a collimator 7, and the collimator 7 couples the spatial light signal into the optical fiber. The collimator 7 is serially connected to an optoelectronic converter 8 to convert the optical signal into an electrical signal. The optoelectronic converter 8 is connected to a low-pass filter 9, and the filter 9 is connected to a digital converter 10. The signal passing through the digital converter 10 is input into the reconstruction model, and the initial reconstruction is first performed, and the result after the initial reconstruction is input into the hybrid reconstruction. In each stage of the hybrid reconstruction, driven by the traditional ISTA, the Transformer backbone is combined with the convolutional layer to iteratively solve the general L1 norm optimization problem. Specifically, ISTA is updated by gradient descent and soft threshold functions, and each iterative layer is optimized according to the current reconstruction result. First, the convolution layer extracts the detailed features of the signal locally, and the encoder further captures the global features of the signal and calculates the adjustment of the error term through gradient descent, and then applies the soft threshold function to sparse the result. The decoder fuses and reconstructs the global features extracted by the encoder with the local information of the convolution layer to generate a more accurate signal. In order to further enhance the reconstructed details, the decoder then adds another convolution layer to refine the output and improve the quality of the local features of the signal, making the final reconstruction result closer to the real signal. The self-attention mechanism of the Transformer helps capture long-range dependencies in each layer iteration, making the final reconstruction result closer to the real signal. Example

[0035] like Figure 1 As shown, a master control terminal and a remote node are taken as an example.

[0036] The laser emits light pulses of different wavelengths into the single-mode optical fiber; the laser is connected to the single-mode optical fiber.

[0037] The laser is a continuous optical signal of 100kHz with a power of 16dBm.

[0038] The single-mode optical fiber is connected to the Mach-Zehnder modulator, which modulates the transmitted signal to obtain optical signals corresponding to different wavelengths.

[0039] The single-mode optical fiber is connected to the Mach-Zehnder modulator to modulate the node signal x(t) to wavelengths λ1 and λ2 respectively. ... λ n In the figure, the DC bias point of the Mach-Zehnder modulator is the quadrature point, and the signal source uses an arbitrary waveform generator to generate a 900MHz sinusoidal RF signal.

[0040] The Mach-Zehnder modulator connects the transmission optical fiber to the receiving control center.

[0041] The transmission fiber length is 20 km and the dispersion φ is 450 ps. 2 .

[0042] The transmission optical fiber is connected to the diffraction grating to disperse and arrange the optical carrier signals of different wavelengths, thereby separating the optical signals of different wavelengths.

[0043] The grating is connected to the spatial light modulator, and discrete sampling of the signal is achieved by controlling the amplitude screen loading rate of the spatial light modulator.

[0044] The grating constant is 1200 lines / mm and the diffraction order is 2.

[0045] The spatial light modulator has a resolution of 1920x1080, a pixel size of 15 microns and a modulation depth of 100%.

[0046] The spatial light modulator is connected to the photoelectric converter to convert the optical signal into an electrical signal, and the photoelectric converter is connected to a low-pass filter for filtering.

[0047] The bandwidth of the photoelectric converter is 40 GHz. The detection signal is recorded by a sampling oscilloscope. The low-pass filter calculates the cumulative sum of the 4-bit mixed signal within the time period, and the compression rate is 4.

[0048] The low-pass filter is connected to the digital converter to convert the analog signal into a digital signal.

[0049] The output signal image is input into the reconstruction network, and the signal is restored through the network model.

[0050] The computing platform is implemented on the platform of PyTorch1.9.0 framework, Intel Xeon Silver4210CPU and GeForceRTX3090GPU. The network training uses the waveform generated by the arbitrary waveform generator and the oscilloscope to collect 2040 waveform images, of which 2000 are used for training and 40 images are used for testing. The mean square error function (MSE) is used as the loss function during training.

[0051] from Figure 2 It can be seen from the results that the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the image reconstructed using compressed sensing based on the autoencoder have obtained relatively good data, with PSNR / SSIM being 37.53 / 0.9684 and 35.07 / 0.9308 respectively; it has a higher reconstruction accuracy.

[0052] This embodiment proposes a spatial optical signal transmission method based on autoencoder compressed sensing. Compared with the prior art solution, the beneficial effects of the present invention are as follows:

[0053] The present invention has a simple structure and a novel reconstruction method. Compared with the existing compressed sensing spatial optical signal processing system, it can effectively solve the problems of high computational cost and limited real-time performance in the reconstruction process. At the same time, the reconstruction method uses the ISTA Transformer backbone network and the CNN deep learning network algorithm signal for reconstruction, which can obtain higher quality reconstruction results under the same compression ratio.

Claims

1. A spatial optical signal transmission method based on self-encoding compressed sensing, characterized in that: The following steps are involved: S1. The transmitted RF signal is amplitude modulated to a multi-wavelength optical carrier through a Mach-Zehnder modulator; S2. The modulated multi-wavelength optical carrier is mixed with a pseudo-random sequence, and random sub-sampling of the signal is achieved by controlling the amplitude screen loading rate of the spatial light modulator; S3. The randomly sub-sampled signal is coupled into the optical fiber through a collimator, and then converted into an electrical signal through a photoelectric converter, and low-pass filtered to obtain a signal waveform image after compression sampling; S4. Use an oscilloscope to collect the original RF signal waveform image, construct a signal data set to input the signal reconstruction network model built based on the autoencoder, output the reconstructed waveform image, and perform pre-training; S5. The waveform images compressed and sampled by S3 are combined into a data set, the pre-trained signal reconstruction network model is fine-tuned, the reconstructed waveform images are output, the spatial optical signal transmission is completed, and the performance is evaluated.

2. The spatial optical signal transmission method based on self-encoding compressed sensing according to claim 1, characterized in that: The specific process of mixing the modulated multi-wavelength optical carrier with the pseudo-random sequence is as follows: the modulated multi-wavelength optical carrier first passes through a diffraction grating to disperse and arrange optical carrier signals of different wavelengths, and then passes through a spatial light modulator to achieve mixing of the signal with the pseudo-random sequence in the frequency domain.

3. The spatial optical signal transmission method based on self-encoding compressed sensing according to claim 2 is characterized in that: The random sub-sampling of the signal is achieved by controlling the loading rate of the amplitude screen of the spatial light modulator. The specific process is: loading a two-dimensional pseudo-random signal with the same size as the spot of the spatial light modulator, controlling the amplitude of the corresponding pixel point of the amplitude screen, and setting the speed of loading the amplitude screen to achieve the random sub-sampling of the signal; The two-dimensional pseudo-random signal is a binary distribution of 0 and 1, corresponding to 0 and 255 pixel amplitudes of the light modulation amplitude, respectively.

4. The spatial optical signal transmission method based on self-encoding compressed sensing according to claim 3 is characterized in that: The implementation process of the signal reconstruction network model is as follows: the input waveform image passes through the initial reconstruction module and the mixed reconstruction module in sequence, and the reconstructed waveform image is output.

5. The spatial optical signal transmission method based on self-encoding compressed sensing according to claim 4 is characterized in that: The initial reconstruction module uses the initial reconstruction matrix to complete the initial reconstruction of the signal; The hybrid reconstruction module is formed by integrating a Transformer backbone network based on an iterative shrinkage threshold algorithm ISTA and a CNN.

6. The spatial optical signal transmission method based on self-encoding compressed sensing according to claim 5, characterized in that: The specific implementation process of the initial reconstruction module is: taking a randomly generated Gaussian matrix as an initial reconstruction matrix, and performing element-by-element multiplication with the input waveform image to obtain an initial reconstructed image.

7. The spatial optical signal transmission method based on self-encoding compressed sensing according to claim 6, characterized in that: The hybrid reconstruction module is specifically implemented as follows: the hybrid reconstruction module is composed of n cascaded restoration stages, and the input initial reconstructed image is passed through the hybrid reconstruction module to generate the final reconstructed photon signal; each stage of the hybrid reconstruction module is composed of a Transformer backbone network and convolutional layers added before and after the Transformer backbone network, and the Transformer backbone network structure includes an encoder and a decoder part, the encoder and the decoder are composed of fully connected layers, and a multi-layer self-attention mechanism is introduced after the encoder and the decoder respectively to capture the local and global features of the image, and a soft threshold function is introduced before the encoder for processing, and reverse training is performed through the iterative shrinkage threshold algorithm ISTA.

Images