Intelligent broadband microwave photon receiving device assisted by deep learning algorithm

Abstract

The invention discloses an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm. The intelligent broadband microwave photon receiving device comprises a microwave photon receiving module and a deep learning processing module, the microwave photon module comprises a receiving antenna, a low-noise amplifier, a distributed feedback semiconductor laser, an optical fiber, an optical amplifier, a photoelectric detector, a low-pass filter and an analog-to-digital converter; the receiving antenna, the low-noise amplifier, the distributed feedback semiconductor laser, the optical fiber, the optical amplifier, the photoelectric detector, the low-pass filter and the analog-to-digital converter are connected in sequence, and the analog-to-digital converter is connected with the deep learning processing module. According to the device, through the combination of the microwave photon receiving module and the deep learning processing module, the dependence on high-performance hardware is reduced, the complexity and cost of the system are reduced, software compensation for hardware performance is realized, and the requirement for expensive and high-precision equipment is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave photonics, and in particular relates to an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm. Background Art

[0002] In the practical application of microwave photonic radar, a photonic analog-to-digital conversion system is generally used as a receiver (Xu S, Liu B, Yi S, et al. "Analog spatiotemporal feature extraction for cognitive radio-frequency sensing with integrated photonics", Light: Science & Applications, 50, 2024). Due to the random scattering of the transmitted signal on the surface of the object, the signal received by the system is usually a broadband signal. In 2020, Zou Weiwen's research group at Shanghai Jiao Tong University proposed and demonstrated an improved deep learning-driven photonic analog-to-digital converter, in which a neural network is used to reduce the signal distortion of the photonic system. This work built an improved U-Net neural network, simulating thousands of radar echo data sets based on the principle of multiple scattering. The trained neural network reduced signal distortion with high quality, indicating that the neural network improved the detection capability of the one-dimensional image of weak targets of the microwave photon radar system (Xu S, Wan J, Wang R, et al. "Modified deep-learning-powered photonic analog-to-digital converter for wideband complicated signal receiving", Optics Letters, 5303-5306, 2020). In 2020, Pan Shilong's research group at Nanjing University of Aeronautics and Astronautics proposed and demonstrated a linear frequency modulation (LFM) signal recovery method based on residual convolutional neural network based on machine learning (Zhou Y, Zhang F, Sun G, et al. "Machinelearning based LFM signal recovery for fiber-connected radar networks", Optoelectronics and Communications Conference, M4B-2, 2021). The experimental results verified the feasibility of the 8-layer CNN network used in broadband radar echo recovery in fiber-connected radar networks.

[0003] Microwave photonics combines the advantages of microwave technology and photonics, and provides solutions for ultra-high-speed, broadband, anti-interference, and long-distance transmission of microwave signals. However, when processing broadband signals in microwave photonics, traditional processing methods face many challenges such as system noise interference, signal attenuation, and distortion caused by nonlinearity of photonic links. In particular, microwave photonic receiving devices are affected by link nonlinearity, and when processing complex broadband signals, clutter signals will appear in the same frequency band, which is difficult to filter out by traditional methods. For modern radar applications, the signal is processed by deep learning after passing through the microwave photonic receiving device, which further improves the system's target detection capability and recognition accuracy, especially when the target features are blurred or the background noise is large; in addition, the system's intelligent processing technology can reduce dependence on high-performance hardware and reduce system complexity and cost.

[0004] The existing scheme relies on an external modulator to achieve broadband signal modulation, which will bring unnecessary noise and nonlinear problems to the receiving system and increase the cost of the device. This scheme does not require an external modulator, and realizes the combination of long-distance transmission of broadband signals and intelligent processing. In the deep learning processing module, in order to learn the complex characteristics of broadband signals, a deep residual convolutional neural network with weighted fusion and jump connections is constructed to achieve high-precision recovery of the distorted signals caused by the microwave photonic system. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm, aiming to solve the problem that the existing digital technology cannot effectively suppress the noise and nonlinear distortion of the microwave photon receiving device.

[0006] To solve the above technical problems, the present invention provides an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm, comprising a microwave photon receiving module and a deep learning processing module; the microwave photon receiving module comprises a receiving antenna, a low-noise amplifier, a distributed feedback semiconductor laser, an optical fiber, an optical amplifier, a photodetector, a low-pass filter and an analog-to-digital converter; the receiving antenna, the low-noise amplifier, the distributed feedback semiconductor laser, the optical fiber, the optical amplifier, the photodetector, the low-pass filter and the analog-to-digital converter are connected in sequence, and the analog-to-digital converter is connected to the deep learning processing module.

[0007] The communication process of the intelligent broadband microwave photon receiving device assisted by the deep learning algorithm includes:

[0008] The receiving antenna receives broadband electromagnetic wave signals, and the output end is connected to a low noise amplifier via a radio frequency cable;

[0009] The output end of the low noise amplifier is connected to the direct modulation port of the distributed feedback semiconductor laser through a radio frequency cable;

[0010] The distributed feedback semiconductor laser has a direct modulation port, and the output end of the distributed feedback semiconductor laser is connected to the optical fiber;

[0011] The optical amplifier input end is connected to the optical fiber, and the optical amplifier output end is connected to the photodetector input end;

[0012] The output end of the photodetector is connected to the input end of the low-pass filter;

[0013] The output end of the low-pass filter is connected to the deep learning processing module;

[0014] The intelligent broadband microwave photon receiving device utilizes the distributed feedback semiconductor laser, the optical fiber and the optical amplifier to realize electro-optical modulation, transmission and amplification of broadband signals;

[0015] The intelligent broadband microwave photon receiving device realizes broadband signal photoelectric conversion, low-frequency filtering and conversion of analog signals to digital signals through the photoelectric detector, the low-pass filter and the analog-to-digital converter;

[0016] The deep learning processing module includes the following steps:

[0017] Step 1, performing short-time Fourier transform on N broadband signals at the output end of the receiving antenna to obtain time-frequency matrix data corresponding to the broadband signal, marking them as clean broadband signal samples, and normalizing the clean broadband signal samples;

[0018] Step 2, performing short-time Fourier transform on the broadband signal received by the microwave photon receiving module, obtaining time-frequency matrix data corresponding to N broadband signals by changing the power of the low-noise amplifier and the length of the optical fiber, marking them as contaminated broadband signal samples, and normalizing the contaminated broadband signal samples;

[0019] Step 3, dividing the N clean broadband signal samples and the N contaminated broadband signal samples into a test set, a validation set and a training set in a ratio of 8:1:1; there is no overlap between the test set, the validation set and the training set;

[0020] Step 4, construct a deep residual convolutional neural network, whose structure is as follows: the first layer, convolution layer + ReLU activation layer; the second to the 16th layers are convolution layer + BN layer + ReLU activation layer; the 17th layer is a convolution layer; the first layer uses Conv+ReLU to realize signal feature extraction, where the convolution kernel size used in the convolution layer is 3×3, and the number of input channels and the number of output channels are 1 and 128 respectively; the second to the 16th layers are composed of the same Conv+BN+ReLU modules, the convolution kernel size used in the convolution layer is 3×3, the number of input channels and the number of output channels are 128 and 128 respectively; the 17th layer is composed of a convolution layer, the convolution kernel size is 3×3, the number of input channels and the number of output channels are 128 and 1 respectively. On the basis of sequentially connecting the first layer and the seventeenth layer, a skip connection is added between the output of the first layer and the output of the sixteenth layer, and the weight is controlled by the weighted fusion coefficient a. The data of the fusion output of the sixteenth layer is D fusion = aD1 + (1-a) D16, where D fusion is the data of the fusion output of the sixteenth layer, D1 is the output data of the first layer, and D16 is the original data output trained sequentially from the first layer to the sixteenth layer;

[0021] Step 5: Use the training set to train the deep residual convolutional neural network, use the adaptive moment estimation (Adam) optimization method to dynamically adjust the learning rate, and iteratively update the parameters of each layer of the convolutional neural network.

[0022] The present invention provides a deep learning algorithm-assisted intelligent broadband microwave photon receiving device, which connects the receiving antenna, the low-noise amplifier, the distributed feedback semiconductor laser, the optical fiber, the optical amplifier, the photodetector, the low-pass filter and the analog-to-digital converter in sequence to form a microwave photon receiving module, which realizes the reception of broadband signals; uses the receiving antenna to receive broadband signals; uses the direct-modulated laser, optical fiber and optical amplifier to realize electro-optical modulation, transmission and amplification of broadband signals; realizes broadband signal photoelectric conversion, low-frequency filtering and conversion of analog signals to digital signals through photodetectors, low-pass filters and analog-to-digital converters; and the deep learning processing module performs adaptive processing on broadband signals to reduce the noise and nonlinear problems generated by the microwave photon receiving module. The present invention combines the microwave photon receiving module and the deep learning processing module to reduce the dependence on high-performance hardware and the demand for high-precision equipment.

[0023] Compared with the prior art, the beneficial effects are:

[0024] This device directly uses distributed feedback semiconductor lasers and optical fiber transmission, without the need for external modulation devices, reducing system complexity and power consumption. External modulation devices are usually large in size, have high driving voltages, and are complex to control, while directly modulating microwave signals through distributed feedback semiconductor lasers is simpler and more stable. This device can transmit and process broadband microwave signals in the optical domain, and then output electrical signals to the deep learning module through an analog-to-digital converter, reducing the processing pressure on the microwave photon module. The jump connection deep learning processing module with weighted fusion performs adaptive processing on broadband signals, reducing the noise and nonlinear problems generated by the microwave photon receiving module, making the signal recovery effect more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a structural schematic diagram of an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm of the present invention;

[0026] Figure 2 A time-frequency diagram of an output signal of a broadband microwave photon receiving device provided in an embodiment;

[0027] Figure 3 A time-frequency diagram of an output signal of an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm provided in an embodiment;

[0028] Figure 4 An autocorrelation function diagram of an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm provided in an embodiment before and after deep learning processing, A is an autocorrelation function diagram of a broadband signal before deep learning processing, and B is an autocorrelation function diagram of a broadband signal after deep learning processing;

[0029] Explanation of the accompanying drawings: 1. Receiving antenna; 2. Low noise amplifier; 3. Distributed feedback semiconductor laser; 4. Optical fiber; 5. Optical amplifier; 6. Photodetector; 7. Low pass filter; 8. Analog-to-digital converter. DETAILED DESCRIPTION

[0030] The specific implementation of an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm of the present invention is described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the intelligent broadband microwave photon receiving device of the present invention includes: a microwave photon receiving module and a deep learning processing module, and the microwave photon receiving module is composed of a receiving antenna 1, a low-noise amplifier 2, a distributed feedback semiconductor laser 3, an optical fiber 4, an optical amplifier 5, a photodetector 6, a low-pass filter 7, and an analog-to-digital converter 8.

[0032] The receiving antenna 1 receives broadband signals, and an output end thereof is connected to an input end of the low noise amplifier 2 via a radio frequency cable.

[0033] Then the broadband signal is directly modulated to the distributed feedback semiconductor laser 3 through the output end of the low noise amplifier 2, and the distributed feedback semiconductor laser 3 is modulated by a small signal to realize the electro-optical conversion of the broadband signal. At the same time, because the distributed feedback semiconductor laser 3 is affected by the modulated signal power and its own noise, noise and nonlinearity are introduced.

[0034] Then the spectrum modulated by the broadband signal is input into the optical amplifier 5 through the optical fiber 4, and the carrier signal and the modulated sideband signal of the distributed feedback semiconductor laser 3 are amplified in the optical domain.

[0035] The output end of the optical amplifier 5 is connected to the photodetector 6 to achieve electro-optical conversion of broadband signals.

[0036] The output end of the photodetector 6 is connected to the input end of the low-pass filter 7; the output end of the low-pass filter 7 is connected to the deep learning processing module, and the broadband signal enters the deep learning processing stage.

[0037] The intelligent broadband microwave photon receiving device utilizes the distributed feedback semiconductor laser 3, the optical fiber 4 and the optical amplifier 5 to realize the electro-optical modulation, transmission and amplification of broadband signals.

[0038] The intelligent broadband microwave photon receiving device realizes broadband signal photoelectric conversion, low-frequency filtering and conversion of analog signals to digital signals through the photoelectric detector 6, the low-pass filter 7 and the analog-to-digital converter 8.

[0039] The deep learning processing module performs deep learning on broadband signals, including the following steps:

[0040] Step 1, performing short-time Fourier transform on the N broadband signals at the output end of the receiving antenna 1 to obtain time-frequency matrix data corresponding to the broadband signals, marking them as clean broadband signal samples, and normalizing the clean broadband signal samples;

[0041] Step 2, performing short-time Fourier transform on the broadband signal received by the microwave photon receiving module, obtaining time-frequency matrix data corresponding to N broadband signals by changing the power of the low-noise amplifier and the length of the optical fiber, marking them as contaminated broadband signal samples, and normalizing the contaminated broadband signal samples;

[0042] Step 3, dividing the N clean broadband signal samples and the N contaminated broadband signal samples into a test set, a validation set and a training set in a ratio of 8:1:1; there is no overlap between the test set, the validation set and the training set;

[0043] Step 4, construct a deep residual convolutional neural network, whose structure is as follows: the first layer, convolution layer + ReLU activation layer; the second to the 16th layers are convolution layer + BN layer + ReLU activation layer; the 17th layer is a convolution layer; the first layer uses Conv+ReLU to realize signal feature extraction, wherein the convolution kernel size used in the convolution layer is 3×3, and the number of input channels and the number of output channels are 1 and 128 respectively; the second to the 16th layers are composed of the same Conv+BN+ReLU module, the convolution kernel size used in the convolution layer is 3×3, the number of input channels and the number of output channels are 128 and 128 respectively; the 17th layer is composed of a convolution layer, the convolution kernel size is 3×3, the number of input channels and the number of output channels are 128 and 1 respectively. On the basis of the sequential connection between the first layer and the 17th layer, a jump connection is added between the output of the first layer and the output of the 16th layer, and the weight is controlled by the weighted fusion coefficient a. The data of the fused output of the 16th layer is D 融合 =aD 1 +(1-a)D 16 , where D 融合 is the data of the 16th layer fusion output, D 1 Output data for layer 1, D 16 is the original data output after training from the 1st layer to the 16th layer in sequence;

[0044] Step 5: Use the training set to train the deep residual convolutional neural network, use an adaptive moment estimation (Adam) optimization method to dynamically adjust the learning rate, and iteratively update the parameters of each layer of the convolutional neural network.

[0045] Example

[0046] The broadband signal 1 is received by the receiving antenna of the microwave photon receiving module and transmitted to the low noise amplifier 2; the distributed feedback semiconductor laser 3, 4 optical fiber and 4 optical amplifier are used to realize the electro-optical modulation, transmission and amplification of the broadband signal; the broadband signal photoelectric conversion, low-frequency filtering and conversion of analog signal to digital signal are realized by the photoelectric detector 6, low-pass filter 7 and analog-to-digital converter 8; and the signal is transmitted to the deep learning processing module for deep learning processing;

[0047] Perform short-time Fourier transform on 500 broadband signals at the output end of the receiving antenna to obtain time-frequency matrix data corresponding to the broadband signal, mark them as clean broadband signal samples, and perform normalization on the clean broadband signal samples;

[0048] The broadband signal received by the microwave photon receiving module is subjected to short-time Fourier transform. By changing the low-noise amplifier amplification power to 5-20 dB and the optical fiber length to 1-10 km, the time-frequency matrix data corresponding to 500 broadband signals are obtained and marked as contaminated broadband signal samples. The contaminated broadband signal samples are normalized. Figure 2The output signal time-frequency diagram of a broadband microwave photon receiving device provided in the embodiment, the radar echo in the simulation experiment is a linear chirp signal with a center frequency of 1.5GHz and a bandwidth of 1.5GHz; after 15dB low-noise amplification, the broadband signal is directly modulated onto a distributed feedback semiconductor laser, transmitted through a 3km single-mode optical fiber, and passed through a 10dB optical amplifier and photodetector; the two-cycle time-frequency curve after low-pass filtering shows that the linear chirp signal produces nonlinear distortion and noise problems after passing through the microwave photon system.

[0049] 500 clean broadband signal samples and 500 contaminated broadband signal samples are divided into a test set, a validation set and a training set in a ratio of 8:1:1; there is no overlap between the test set, the validation set and the training set;

[0050] A deep residual convolutional neural network is constructed, and its structure is as follows: the first layer, convolution layer + ReLU activation layer; the second to the 16th layers are convolution layer + BN layer + ReLU activation layer; the 17th layer is a convolution layer; the first layer uses Conv+ReLU to realize signal feature extraction, and the convolution kernel size used in the convolution layer is 3×3, and the number of input channels and the number of output channels are 1 and 128 respectively; the second to the 16th layers are composed of the same Conv+BN+ReLU modules, the convolution kernel size used in the convolution layer is 3×3, and the number of input channels and the number of output channels are 128 and 128 respectively; the 17th layer is composed of a convolution layer, the convolution kernel size is 3×3, and the number of input channels and the number of output channels are 128 and 1 respectively. On the basis of sequentially connecting the first layer and the seventeenth layer, a skip connection is added between the output of the first layer and the output of the sixteenth layer, and the weight is controlled by the weighted fusion coefficient a. The data of the fusion output of the sixteenth layer is D fusion = aD1 + (1-a) D16, where D fusion is the data of the fusion output of the sixteenth layer, D1 is the output data of the first layer, and D16 is the original data output trained sequentially from the first layer to the sixteenth layer. In this example, a = 0.2;

[0051] The deep residual convolutional neural network is trained with the training set, an adaptive moment estimation (Adam) optimization method is adopted, a learning rate is 0.001, the learning rate is dynamically adjusted, and the parameters of each layer of the convolutional neural network are iteratively updated;

[0052] Reference Figure 3 As shown, the output signal time-frequency diagram of a deep learning algorithm-assisted intelligent broadband microwave photon receiving device provided in the embodiment is compared with the reference Figure 2, in the same power range, after 200 iterations of training of the jump connection deep learning network with weighted fusion, the test set data was used for testing. The clutter signal was effectively removed, and the distortion and noise problems of the microwave photon receiving system were improved. In order to compare the pulse compression ability of the waveform before and after deep learning processing, the time-frequency signal was inverse fast Fourier transformed, and the autocorrelation function of the linear chirp signal was calculated after obtaining the corresponding time domain signal. Figure 4 As shown, its peak-to-base ratio (PFR) is improved by 26.6dB. The addition of the intelligent processing module can achieve high-precision recovery of the distorted signal caused by the microwave photonic system.

[0053] This embodiment uses a directly modulated laser, optical fiber and optical amplifier to realize electro-optical modulation, transmission and amplification of broadband signals; through photoelectric detectors, low-pass filters and analog-to-digital converters, broadband signal optoelectronic conversion, low-frequency filtering and conversion of analog signals to digital signals are realized; the jump connection deep learning processing module with weighted fusion performs adaptive processing on broadband signals, and the jump connection with weighted fusion helps the network to better retain image details, reduce detail loss problems, reduce noise and nonlinear problems generated by the microwave photon receiving module, and restore the broadband signal with high precision.

[0054] The present invention discloses an intelligent broadband microwave photon receiving device assisted by a deep learning algorithm, and the feasibility of the device is verified by simulation. First, a microwave photon receiving module is constructed and simulated using OptiSystem software, where the global variable is set to 10μs and the number of sampling points is 1048576. Here we demonstrate that a linear chirp signal with a center frequency of 1.5GHz and a bandwidth of 1.5GHz is amplified by an amplifier and then input into a directly modulated DFB laser.

[0055] The experimental environment of the deep learning processing module is configured as follows: The processor uses the Intel(R) Core(TM) i9-10900 with a 10-core 20-thread architecture, which can provide powerful computing power and multi-threaded processing capabilities, and can effectively accelerate the training process of deep learning models. CUDA-11.6 is a parallel computing platform and programming model provided by NVIDIA. It can use the parallel computing power of NVIDIA GPU to accelerate the execution of deep learning tasks. CUDA-11.6 has good compatibility with NVIDIA RTX 4090 graphics card and can give full play to the computing potential of the graphics card. CUDNN-8.4 is a deep learning library provided by NVIDIA, which provides high-performance GPU acceleration for deep learning frameworks. PyTorch is compatible with CUDA-11.6 and CUDNN-8.4. Through the above environment, the computing power of GPU can be fully utilized to improve the efficiency of deep learning training, and a high level of deep learning training performance can be obtained, which can effectively accelerate the training process of deep learning models.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the present invention. A person skilled in the art should understand that the specific implementation of the present invention can be modified or replaced by equivalents, but these modifications or changes are within the scope of protection of the pending claims.

Claims

1. A deep learning algorithm-assisted intelligent broadband microwave photon receiving device, characterized in that: include: Microwave photon receiving module and deep learning processing module; The microwave photon receiving module comprises a receiving antenna, a low noise amplifier, a distributed feedback semiconductor laser, an optical fiber, an optical amplifier, a photodetector, a low pass filter and an analog-to-digital converter; the receiving antenna, the low noise amplifier, the distributed feedback semiconductor laser, the optical fiber, the optical amplifier, the photodetector, the low pass filter and the analog-to-digital converter are connected in sequence to realize the reception of broadband signals; The analog-to-digital converter is connected to the deep learning processing module.

2. According to claim 1, a deep learning algorithm-assisted intelligent broadband microwave photon receiving device is characterized in that: The receiving antenna receives a broadband signal, and an output end is connected to an input end of the low noise amplifier through a radio frequency cable.

3. According to claim 2, a deep learning algorithm-assisted intelligent broadband microwave photon receiving device is characterized in that: The output end of the low noise amplifier is connected to the direct modulation port of the distributed feedback semiconductor laser through a radio frequency cable, and the broadband signal is directly modulated to the distributed feedback semiconductor laser through the output end of the low noise amplifier, and the distributed feedback semiconductor laser is modulated by a small signal to realize the electro-optical conversion of the broadband signal.

4. According to claim 3, a deep learning algorithm-assisted intelligent broadband microwave photon receiving device is characterized in that: The modulated broadband signal spectrum is input into the optical amplifier through the optical fiber, and the carrier frequency signal and the modulated sideband signal of the distributed feedback semiconductor laser are amplified in the optical domain.

5. According to claim 4, a deep learning algorithm-assisted intelligent broadband microwave photon receiving device is characterized in that: The output end of the optical amplifier is connected to the photoelectric detector to achieve electro-optical conversion of broadband signals.

6. The deep learning algorithm-assisted intelligent broadband microwave photon receiving device according to claim 5, characterized in that: The output end of the photodetector is connected to the input end of the low-pass filter; the output end of the low-pass filter is connected to the deep learning processing module, and the broadband signal enters the deep learning processing stage.

7. The deep learning algorithm-assisted intelligent broadband microwave photon receiving device according to claim 6, characterized in that: The intelligent broadband microwave photon receiving device utilizes the distributed feedback semiconductor laser, the optical fiber and the optical amplifier to realize the electro-optical modulation, transmission and amplification of broadband signals.

8. The deep learning algorithm-assisted intelligent broadband microwave photon receiving device according to claim 7, characterized in that: The intelligent broadband microwave photon receiving device realizes broadband signal photoelectric conversion, low-frequency filtering and conversion of analog signals to digital signals through the photoelectric detector, the low-pass filter and the analog-to-digital converter.

9. The deep learning algorithm-assisted intelligent broadband microwave photon receiving device according to claim 1, characterized in that: The deep learning processing module performs deep learning on the broadband signal, comprising the following steps: Step 1, performing short-time Fourier transform on the N broadband signals at the output end of the receiving antenna to obtain time-frequency matrix data corresponding to the broadband signals, marking them as clean broadband signal samples, and normalizing the clean broadband signal samples; Step 2, performing short-time Fourier transform on the broadband signal received by the microwave photon receiving module, obtaining time-frequency matrix data corresponding to N broadband signals by changing the power of the low-noise amplifier and the length of the optical fiber, marking them as contaminated broadband signal samples, and normalizing the contaminated broadband signal samples; Step 3, dividing the N clean broadband signal samples and the N contaminated broadband signal samples into a test set, a validation set and a training set in a ratio of 8:1:1; there is no overlap between the test set, the validation set and the training set; Step 4, construct a deep residual convolutional neural network, whose structure is as follows: the first layer, convolution layer + ReLU activation layer; the second to the 16th layers are convolution layer + BN layer + ReLU activation layer; the 17th layer is a convolution layer; the first layer uses Conv+ReLU to realize signal feature extraction, wherein the convolution kernel size used in the convolution layer is 3×3, and the number of input channels and the number of output channels are 1 and 128 respectively; the second to the 16th layers are composed of the same Conv+BN+ReLU module, the convolution kernel size used in the convolution layer is 3×3, the number of input channels and the number of output channels are 128 and 128 respectively; the 17th layer is composed of a convolution layer, the convolution kernel size is 3×3, the number of input channels and the number of output channels are 128 and 1 respectively; on the basis of the sequential connection between the first layer and the 17th layer, a jump connection is added between the output of the first layer and the output of the 16th layer, and the weight is controlled by the weighted fusion coefficient a, and the data of the fused output of the 16th layer is D 融合 =aD1+(1-a)D 16 , where D 融合 is the fusion output data of the 16th layer, D1 is the output data of the 1st layer, and D 16 is the original data output after training from the 1st layer to the 16th layer in sequence; Step 5: Use the training set to train the deep residual convolutional neural network, use an adaptive moment estimation (Adam) optimization method to dynamically adjust the learning rate, and iteratively update the parameters of each layer of the convolutional neural network.

Images