Optical communication network adaptive entropy loading method and device based on transfer learning

Abstract

The invention discloses an optical communication network adaptive entropy loading method and device based on transfer learning, and belongs to the technical field of optical communication, and the method specifically comprises the steps of entropy loading model building, synthetic signal generation and model pre-training, experimental signal generation and model transfer learning, and adaptive entropy loading. In the whole process, the entropy distribution process of traditional entropy loading is replaced by the neural network, and the entropy approaching the channel capacity is predicted according to the received signal. A large number of synthetic signals are used for pre-training an entropy loading model, transfer learning is combined, the entropy loading model predicts the entropy of an unknown channel in the optical communication network in a self-adaptive mode, dependence of an existing entropy loading algorithm on the channel state is reduced, and therefore self-adaptive entropy loading of the optical communication network is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical communication, and particularly relates to an optical communication network adaptive entropy loading method and device based on transfer learning. Background Art

[0002] As an important part of modern communication networks, optical communication technology carries a large number of communication services. With the development of emerging digital information services such as cloud computing, high-definition streaming media, and generative artificial intelligence, and the popularization of various mobile devices, optical communication faces more complex and huge data traffic, posing higher requirements for the resource allocation and expansion of optical communication networks in terms of transmission capacity, transmission rate, and dynamic characteristics. At the same time, due to the congestion of existing communication frequency bands and the gradual concentration of available spectrum resources towards high frequencies, optical communication systems and network devices cause more significant frequency-selective fading to high-frequency signals, resulting in a decline in transmission performance and an inability to fully utilize limited bandwidth resources.

[0003] Adaptive loading technology analyzes the characteristics of the communication channel and adjusts the signal configuration according to the channel state when transmitting signals, which can maximize the utilization of spectrum resources and is one of the important methods to improve the communication rate. Classic adaptive loading algorithms include the Water Filling algorithm that can theoretically reach the Shannon limit and the Bit and Power Loading (BPL) algorithm that mainly allocates modulation formats and signal power, such as the Chow algorithm, Fischer algorithm, and Levin-Campello algorithm, etc., which have been widely used in optical communication. In recent years, due to the emergence of Probabilistic Constellation Shaping (PCS) technology, it has become possible for the source entropy to follow a Gaussian distribution, and the adaptive loading technology closer to the channel capacity has received attention and research. Through literature retrieval, it is found that the paper "Entropy-Loading: The Multi-Carrier Constellation-Shaping for Colored-SNR Optical Channels" proposed the Entropy Loading algorithm based on the Water Filling algorithm and PCS technology, achieving a capacity improvement of the Orthogonal Frequency Division Multiplexing (OFDM) signal closer to the Shannon limit. Similarly, the paper "Uniform Entropy Loading for Pre-coded DMT Systems in Fading Optical Channel" proposed a uniform entropy loading algorithm for pre-equalizing the Signal-Noise Ratio (SNR), which achieved better network data rate and peak-to-average power ratio for Discrete Multi-Tone (DMT) signals compared with the LC algorithm. The paper "Entropy Allocation Optimization for PS-OFDM With Constellation Partitioning Based Modeling" proposed an entropy allocation scheme optimized based on the projected mirror descent algorithm, which allocated more entropy to high-SNR channels compared with the entropy allocation based on the Shannon limit.

[0004] In summary, the existing entropy loading algorithms need to predict the SNR of the sub-channels in the optical communication system to achieve entropy allocation. However, the channels in the optical communication network change more dynamically and complexly. The above-mentioned entropy loading algorithms are dependent on the channel state, which limits their application in dynamic networks. Therefore, it is necessary to provide an adaptive entropy loading scheme that can be applied to optical communication networks with unknown channel states. Summary of the Invention

[0005] In view of the above, the object of the present invention is to provide an adaptive entropy loading method and device for an optical communication network based on transfer learning, which replaces the entropy allocation process of traditional entropy loading with an entropy loading model, and predicts the entropy approaching the channel capacity according to the received signal. A large number of synthetic signals are used to pre-train the entropy loading model, combined with transfer learning, so that the entropy loading model can adaptively predict the entropy of the unknown channel response in the optical communication network, reduce the dependence of the existing entropy loading algorithm on the channel state, and thus achieve adaptive entropy loading of the optical communication network.

[0006] To achieve the above object of the invention, an adaptive entropy loading method for an optical communication network based on transfer learning provided by an embodiment includes the following steps: Building the entropy loading model: Build an entropy loading model including an entropy prediction module and a probability shaping module. Among them, the entropy prediction module is used to predict the entropy approaching the channel capacity based on signal features, and the probability shaping module is used to perform entropy allocation for each sub-carrier using probability amplitude shaping technology to complete the entropy loading of the signal; Generating synthetic signals and pre-training the model: Generate synthetic signals by simulating channel characteristics, and use the synthetic signals to pre-train the entropy loading model to optimize the model parameters; Generating experimental signals and performing model transfer learning: Collect experimental signals through optical communication network experiments, and use the experimental signals to perform transfer learning on the pre-trained entropy loading model to achieve parameter fine-tuning; Adaptive entropy loading: Use the entropy loading model with fine-tuned parameters to predict the entropy of the optical communication network signal with unknown channel response, and use probability amplitude shaping technology to perform entropy allocation for each sub-carrier to complete the entropy loading of the signal.

[0007] Preferably, the entropy prediction module uses a deep neural network.

[0008] Preferably, generating synthetic signals by simulating channel characteristics includes: Converting the serial bit sequence into a parallel bit sequence, performing quadrature amplitude modulation symbol mapping on the parallel bit sequence to obtain a symbol sequence, generating a discrete multi-tone modulation signal based on the inverse fast Fourier transform of the symbol sequence, inserting a cyclic prefix, and then converting it into a serial sequence for simulating channel transmission; Filtering the additive white Gaussian noise channel to simulate the frequency-selective response, and simulating the transmission of the discrete multi-tone modulation signal to obtain a spectrally non-flat received signal; Perform serial-to-parallel conversion on the received signal, remove the cyclic prefix of the parallel sequence, and obtain quadrature amplitude modulation symbols through fast Fourier transform. Combine the water-filling algorithm to calculate the entropy corresponding to each subcarrier. Use the in-phase and quadrature components of the quadrature amplitude modulation symbols and the average SNR of the received signal as signal features, and use the entropy as a feature label to form a composite signal.

[0009] Preferably, use the composite signal to pre-train and optimize the model parameters of the entropy loading model, including: Input the signal features in the composite signal into the entropy loading model to predict the entropy approaching the channel capacity. Use the smooth-L1 function between the entropy prediction value and the feature label as the loss function, and calculate the gradient and backpropagation to update the network parameters of the entropy prediction module in the entropy loading model to achieve parameter optimization of the model.

[0010] Preferably, collect experimental signals through optical communication network experiments, including: Simulate the dynamic changes of the optical communication network through experiments of transmitting discrete multi-tone modulation signals with different frequencies under different link conditions. After synchronizing, channel equalizing, and performing fast Fourier transform on the experimental received signals in sequence, use the in-phase and quadrature components of the quadrature amplitude modulation symbols and the average SNR of the experimental received signal as signal features, and use the entropy approaching the capacity as a feature label to form an experimental signal.

[0011] Preferably, perform transfer learning on the pre-trained entropy loading model using the experimental signal to achieve parameter fine-tuning, including: Freeze the network layer parameters of the entropy prediction module in the pre-trained entropy loading model, and add a part of trainable network layers. On this basis, use the experimental signal to fine-tune the newly added network layers to achieve transfer learning.

[0012] Preferably, use the entropy loading model with fine-tuned parameters to predict the entropy of the optical communication network signal with unknown channel response, including: Input the signal features of the optical communication network signal with unknown channel response into the entropy prediction module of the entropy loading model to predict the entropy of the signal.

[0013] To achieve the above invention purpose, an embodiment of the present invention also provides an optical communication network adaptive entropy loading device based on transfer learning, including: An entropy loading model building unit, which is used to build an entropy loading model including an entropy prediction module and a probability shaping module. Among them, the entropy prediction module is used to predict the entropy approaching the channel capacity based on signal features, and the probability shaping module is used to perform entropy allocation for each subcarrier using probability amplitude shaping technology to complete the entropy loading of the signal; A synthetic signal generation and model pre-training unit, which is used to generate synthetic signals by simulating channel characteristics and pre-train the entropy loading model using the synthetic signals to optimize the model parameters; An experimental signal generation and model transfer learning unit: which is used to collect experimental signals through optical communication network experiments and perform transfer learning on the pre-trained entropy loading model using the experimental signals to achieve parameter fine-tuning; An adaptive entropy loading unit, which uses the entropy loading model after parameter fine-tuning to predict the entropy of signals with unknown channel responses, and uses probability amplitude shaping technology to perform entropy allocation for each sub-carrier to complete the entropy loading of the signals.

[0014] To achieve the above invention purpose, the embodiment also provides a computing device, including a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the above-mentioned adaptive entropy loading method for optical communication networks based on transfer learning.

[0015] To achieve the above invention purpose, the embodiment also provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the above-mentioned adaptive entropy loading method for optical communication networks based on transfer learning.

[0016] Compared with the prior art, the beneficial effects of the present invention at least include: The present invention constructs an entropy loading model including an entropy prediction module and a probability shaping module. First, it pre-trains through synthetic signals to optimize the model parameters, then performs transfer learning on the pre-trained model based on experimental signals simulated from the optical communication network to adapt to the signal characteristics of the optical communication network. Finally, it uses the entropy loading model after transfer learning to predict the entropy of optical communication network signals with unknown channel responses, and uses probability amplitude shaping technology to perform entropy allocation for each sub-carrier to complete the entropy loading of the signals. The entire process does not depend on the channel state of the network, realizes the adaptive entropy loading of optical communication networks with unknown channel states, and meets the application requirements in dynamic optical communication networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a flowchart of the adaptive entropy loading method for optical communication networks based on transfer learning provided by the embodiment; Figure 2It is a schematic structural diagram of the entropy loading model provided by the embodiment; Figure 3 It is a generation process diagram of the synthetic signal provided by the embodiment; Figure 4 It is a pre-training process diagram of the divine entropy loading model provided by the embodiment; Figure 5 It is a generation process diagram of the experimental signal provided by the embodiment; Figure 6 It is a transfer learning process diagram provided by the embodiment; Figure 7 It is a schematic structural diagram of the optical communication network adaptive entropy loading device based on transfer learning provided by the embodiment. Detailed implementation manners

[0019] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation manners described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0020] The technical concept of the present invention is as follows: In the existing entropy loading allocation process, it is necessary to accurately measure the SNR of each sub-channel in the communication system in advance and then perform entropy allocation. In this way, for an optical communication network with an unknown channel response state (that is, the SNR of each sub-channel is unknown), the existing entropy loading allocation process is not applicable. For this reason, the embodiments of the present invention provide an optical communication network adaptive entropy loading method based on transfer learning, which first constructs a mapping relationship between the signal features of all channels and entropy loading allocation under different channel response states in a new optical communication network through pre-training and transfer learning, where the average SNR included in the signal features is used as basic label knowledge. In this way, based on this mapping relationship for unknown channels, it is not necessary to re-measure and calculate the accurate SNR of each sub-channel, and entropy loading allocation can be performed based on this mapping relationship.

[0021] As Figure 1 shown, the embodiment provides an optical communication network adaptive entropy loading method based on transfer learning, including the following steps: S1, Entropy loading model construction: Construct an entropy loading model including an entropy prediction module and a probability shaping module, where the entropy prediction module is used to predict the entropy approaching the channel capacity based on signal features, and the probability shaping module is used to perform entropy allocation for each sub-carrier using probability amplitude shaping technology to complete the entropy loading of the signal.

[0022] In the embodiment, the entropy prediction module uses a deep neural network, which includes an input layer, a hidden layer, and an output layer. Specifically, it can be as Figure 2As shown, it includes 1 input layer, 4 hidden layers, and 1 output layer, all of which are fully connected layers. The input layer is the in-phase and quadrature components of each subcarrier of the Discrete Multi-Tone (DMT) signal and the average SNR of the DMT signal, and the output layer is the entropy predicted by the deep neural network. The probability shaping module performs probability shaping and entropy allocation on each subcarrier of the DMT signal according to the predicted entropy and the Probabilistic Amplitude Shaping (PAS) technique to complete the entropy loading of the synthesized signal.

[0023] S2, Synthesized signal generation and model pre-training: Generate a synthesized signal by simulating the channel characteristics, and use the synthesized signal to pre-train the entropy loading model to optimize the model parameters.

[0024] In the embodiment, a synthesized signal is generated by simulating the channel characteristics as a training sample, as Figure 3 shown. The specific process is as follows: First, DMI signal generation: Convert the serial bit sequence into a parallel bit sequence, perform quadrature amplitude modulation symbol mapping on the parallel bit sequence to obtain a symbol sequence, generate a DMT signal based on the inverse fast Fourier transform for the symbol sequence, insert a cyclic prefix, and then convert it into a serial sequence for simulating channel transmission; Then, frequency-selective response channel simulation: Filter the Additive White Gaussian Noise (AWGN) channel to simulate the frequency-selective response (for example, select a Butterworth filter for low-pass filtering to simulate the frequency-selective fading response), and simulate the transmission of the DMT signal to obtain a received signal with uneven spectrum; Finally, synthesized signal acquisition: Perform serial-to-parallel conversion on the received DMT signal, remove the cyclic prefix of the parallel sequence, and obtain M-order quadrature amplitude modulation symbols through the fast Fourier transform. Combine the entropy corresponding to each subcarrier calculated according to the water-filling algorithm approaching the Shannon limit. Use the in-phase and quadrature components of the M-order quadrature amplitude modulation symbols and the average SNR of the received signal as signal features, and use the entropy as a feature label to form a synthesized signal.

[0025] In the embodiment, use the synthesized signal to pre-train the entropy loading model to optimize the model parameters, including: Input the signal features in the synthesized signal into the entropy loading model to predict the entropy approaching the channel capacity. Use the smooth-L1 function between the entropy prediction value and the feature label as the loss function, and calculate the gradient and backpropagation to update the network parameters of the entropy prediction module in the entropy loading model to achieve parameter optimization of the model.

[0026] Specifically, as Figure 4 shown, the pre-training process is as follows: First, randomly divide the pre-training sample set into a training set and a validation set according to a ratio of 8 to 2. For forward propagation, use the rectified linear unit function as the activation function for the hidden layer and the linear function as the activation function for the output layer. Obtain the predicted value through the activation function and compare it with the feature labels of the training set; Then, for gradient calculation, use the smooth-L1 function as the loss function. Calculate the loss value obtained from the entropy prediction value and the feature labels, which reflects the result of forward propagation. Take the derivative of the loss value with respect to the weights for each neuron in each layer of the model to obtain the gradient of each neuron for model weight optimization; Next, for backpropagation, use the Adam optimizer to update the gradient, set an appropriate learning rate to start training. For example, set the learning rate to 0.0001 to start training; Finally, during the training process, perform batch normalization and Dropout strategies before and after the activation function respectively to prevent gradient explosion and overfitting problems. After multiple rounds of training (for example, 500 rounds), obtain the optimal neural network model parameters.

[0027] S3. Experimental signal generation and model transfer learning: Collect experimental signals through optical communication network experiments and use the experimental signals for transfer learning of the pre-trained entropy loading model to achieve parameter fine-tuning.

[0028] In the embodiment, collect experimental signals as a small amount of sample data through optical communication network experiments for transfer learning of the entropy loading model. Among them, the process of collecting experimental signals is as follows: Simulate the dynamic changes of the optical communication network through experiments of transmitting discrete multi-tone modulation signals with different frequencies under different link conditions. After synchronizing, channel equalizing, and performing fast Fourier transform on the experimental received signals in sequence, use the in-phase and quadrature components of the quadrature amplitude modulation symbols and the average SNR of the experimental received signals as signal features, and use the capacity-approaching entropy as the feature label to form experimental signals.

[0029] The transmitting end sends a multi-carrier signal with a sub-carrier normalized signal power of S 0 The receiving end measures the signal-to-noise ratio of SNR i sub-carriers based on the received signal, calculates the power of each sub-carrier S i The receiving end calculates the channel capacity of each sub-carrier, that is, the capacity-approaching entropy of each sub-carrier, according to the power of each sub-carrier S i as: H i where: ; where, in the formula mean (·) represents taking the average.

[0030] As Figure 5 shown, the optical communication network experimental system includes a signal generation part, a signal transmission part, a signal transmission part, a signal reception part, and a signal processing part. The signal generation part is used to generate a DMT signal and convert it into an analog signal; the signal transmission part is used to modulate the DMT signal output by the signal generation part onto an optical carrier to form an optical signal and transmit it; the signal transmission part is used to transmit the DMT optical signal output by the signal transmission part under different channel conditions; the signal reception part is used to receive the transmitted DMT optical signal and convert it into a digital signal; the signal processing part is used to recover the DMT signal samples and perform adaptive entropy loading.

[0031] As Figure 5 shown, the signal generation part includes a DMT signal generation module and an arbitrary waveform generator module connected in sequence; the signal transmission part includes a laser, a polarization controller, and an electro-optic modulator connected in sequence; the signal transmission part includes fiber back-to-back transmission or fiber transmission; the signal reception wavelength division includes an adjustable optical attenuator and a photodetector connected in sequence; the signal processing part includes a digital storage oscilloscope module, a synchronization module, a channel equalization module, a DMT demodulation module, a feature and label acquisition module connected in sequence. The DMT signal generation module is used to generate DMT signals of different frequencies without entropy loading, and the arbitrary waveform generator module is used to convert the digital DMT signal into an analog signal and input it into the electro-optic modulator. The laser outputs an optical carrier, which enters the electro-optic modulator through the polarization controller. The polarization controller is used to adjust the polarization state of the input optical signal to maximize the output power, and the electro-optic modulator is used to modulate the analog DMT signal generated by the signal generation part onto the optical carrier to form an optical signal and output it. The signal transmission part is used to transmit the optical signal output by the signal transmission part under different channel conditions, such as back-to-back transmission or fiber transmission, and the optical amplifier in the fiber transmission channel is used to increase the signal power after fiber transmission. The adjustable optical attenuator is used to adjust the optical power of the optical signal output by the signal transmission part, and the photodetector converts the optical signal output by the adjustable optical attenuator into an electrical signal for input to the signal processing part. The digital storage oscilloscope converts the analog electrical signal output by the signal reception part into a digital signal for subsequent digital signal processing. The synchronization module can complete synchronization by using the method of pseudo-random noise code, and the channel equalization part can estimate the channel response by using the method of inserting pilots. The DMT signal demodulation module is used to perform serial-to-parallel conversion, remove the cyclic prefix, and perform fast Fourier transform on the signal data after channel equalization in sequence. The in-phase quadrature component and the average SNR of the received signal are collected as features, the entropy approaching the capacity is calculated as the feature label, and the experimental signal is constructed.

[0032] In the embodiment, the specific process of using the experimental signal to perform transfer learning on the pre-trained entropy loading model to adapt to different channel conditions and achieve parameter fine-tuning is as Figure 6As shown, for the entropy loading model pre-trained based on a large number of synthetic signals, the best entropy loading model parameters after pre-training are saved. A small number of experimental signals with different frequency responses are selected for transfer learning. By freezing the network layer parameters of the entropy prediction module in the pre-trained entropy loading model and adding a part of trainable network layers, specifically including: by freezing the original network hidden layer parameters and adding a new trainable hidden layer, fine-tuning the newly added network structure on the basis of the best model parameters to adapt to the experimental signal characteristics, for predicting the entropy of experimental signals with unknown frequency responses.

[0033] S4, Adaptive entropy loading: Using the entropy loading model with fine-tuned parameters to predict the entropy of optical communication network signals with unknown channel responses, and adopting probability amplitude shaping technology to perform entropy allocation for each sub-carrier to complete the entropy loading of the signal.

[0034] In the embodiment, the entropy loading model after transfer learning can be used to perform entropy loading on optical communication network signals with unknown channel responses. Specifically, the signal characteristics of optical communication network signals with unknown channel responses are collected and input into the entropy loading model for entropy prediction. Based on the PAS technology, entropy allocation is performed for each sub-carrier of the signal to complete adaptive entropy loading and optimize the spectral efficiency of the optical communication network.

[0035] As Figure 7 shown, the optical communication network adaptive entropy loading device provided by the embodiment includes: an entropy loading model building unit, a synthetic signal generation and model pre-training unit, an experimental signal generation and model transfer learning unit, and an adaptive entropy loading unit. Among them, the entropy loading model building unit is used to build an entropy loading model including an entropy prediction module and a probability shaping module; the synthetic signal generation and model pre-training unit is used to generate synthetic signals by simulating channel characteristics and pre-train the entropy loading model with the synthetic signals to optimize the model parameters; the experimental signal generation and model transfer learning unit collects experimental signals through optical communication network experiments and performs transfer learning on the pre-trained entropy loading model to achieve parameter fine-tuning; the adaptive entropy loading unit uses the entropy loading model with fine-tuned parameters to predict the entropy of signals with unknown channel responses, and adopts probability amplitude shaping technology to perform entropy allocation for each sub-carrier to complete the entropy loading of the signal.

[0036] It should be noted that when the above-described optical communication network adaptive entropy loading device based on transfer learning performs optical communication network adaptive entropy loading, the above-described division of each functional module should be used as an example. The above functions can be allocated to different functional modules according to needs, that is, the internal structure of the terminal or server is divided into different functional modules or units to complete all or part of the functions described above. In addition, the above-described optical communication network adaptive entropy loading device based on transfer learning and the embodiment of the optical communication network adaptive entropy loading method based on transfer learning belong to the same concept. For the specific implementation process, please refer to the embodiment of the optical communication network adaptive entropy loading method based on transfer learning, which will not be elaborated here.

[0037] Based on the same inventive concept, the embodiment also provides a computing device, including a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the above-described optical communication network adaptive entropy loading method based on transfer learning, which specifically includes the following steps: S1, Entropy loading model construction: Construct an entropy loading model including an entropy prediction module and a probability shaping module. Among them, the entropy prediction module is used to predict the entropy approaching the channel capacity based on signal characteristics, and the probability shaping module is used to perform entropy allocation for each subcarrier using probability amplitude shaping technology to complete the entropy loading of the signal; S2, Synthetic signal generation and model pre-training: Generate a synthetic signal by simulating channel characteristics, and use the synthetic signal to pre-train the entropy loading model to optimize the model parameters; S3, Experimental signal generation and model transfer learning: Collect experimental signals through optical communication network experiments, and use the experimental signals to perform transfer learning on the pre-trained entropy loading model to achieve parameter fine-tuning; S4 Adaptive entropy loading: Use the entropy loading model with fine-tuned parameters to predict the entropy of the optical communication network signal with unknown channel response, and use probability amplitude shaping technology to perform entropy allocation for each subcarrier to complete the entropy loading of the signal.

[0038] For the computing device provided by the embodiment, at the hardware level, in addition to including a processor and a memory, it also includes an internal bus, a network interface, a memory, and other hardware required for other services. The memory is a non-volatile memory. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above-described optical communication network adaptive entropy loading method based on transfer learning in S1-S4. Of course, in addition to the software implementation method, the present invention does not exclude other implementation methods, such as logical devices or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logical unit, and can also be hardware or logical devices.

[0039] Based on the same inventive concept, the embodiment also provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, the above-mentioned optical communication network adaptive entropy loading method based on transfer learning is implemented, which specifically includes the following steps: S1, Entropy loading model construction: Construct an entropy loading model including an entropy prediction module and a probability shaping module. Among them, the entropy prediction module is used to predict the entropy approaching the channel capacity based on signal characteristics, and the probability shaping module is used to perform entropy allocation for each subcarrier by using probability amplitude shaping technology to complete the entropy loading of the signal; S2, Synthetic signal generation and model pre-training: Generate a synthetic signal through simulating channel characteristics, and use the synthetic signal to pre-train the entropy loading model to optimize the model parameters; S3, Experimental signal generation and model transfer learning: Collect experimental signals through optical communication network experiments, and use the experimental signals to perform transfer learning on the pre-trained entropy loading model to achieve parameter fine-tuning; S4 Adaptive entropy loading: Use the entropy loading model with fine-tuned parameters to predict the entropy of the optical communication network signal with unknown channel response, and use probability amplitude shaping technology to perform entropy allocation for each subcarrier to complete the entropy loading of the signal.

[0040] In the embodiment, the computer-readable medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data.

[0041] The above-mentioned specific embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, supplements, equivalent replacements, etc. made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive entropy loading method for an optical communication network based on transfer learning, characterized in that, It includes the following steps: Entropy loading model construction: Construct an entropy loading model including an entropy prediction module and a probability shaping module. Among them, the entropy prediction module is used to predict the entropy approaching the channel capacity based on signal features, and the probability shaping module is used to perform entropy allocation for each subcarrier using probability amplitude shaping technology to complete the entropy loading of the signal; Synthetic signal generation and model pre-training: Generate a synthetic signal through simulating channel characteristics, and use the synthetic signal to pre-train the entropy loading model to optimize the model parameters; Experimental signal generation and model transfer learning: Collect experimental signals through optical communication network experiments, and use the experimental signals to perform transfer learning on the pre-trained entropy loading model to achieve parameter fine-tuning; Adaptive entropy loading: Use the entropy loading model with fine-tuned parameters to predict the entropy of the optical communication network signal with unknown channel response, and perform entropy allocation for each subcarrier using probability amplitude shaping technology to complete the entropy loading of the signal.

2. The adaptive entropy loading method for an optical communication network based on transfer learning according to claim 1, wherein The entropy prediction module uses a deep neural network.

3. The adaptive entropy loading method for an optical communication network based on transfer learning according to claim 1, characterized in that Generating a synthetic signal through simulating channel characteristics includes: Converting the serial bit sequence into a parallel bit sequence, performing quadrature amplitude modulation symbol mapping on the parallel bit sequence to obtain a symbol sequence, generating a discrete multi-tone modulation signal based on the inverse fast Fourier transform for the symbol sequence, inserting a cyclic prefix, and then converting it into a serial sequence for simulating channel transmission; Filtering the additive white Gaussian noise channel to simulate the frequency selective response, and simulating the transmission of the discrete multi-tone modulation signal to obtain a received signal with uneven spectrum; Performing serial-to-parallel conversion on the received signal, removing the cyclic prefix of the parallel sequence, and obtaining quadrature amplitude modulation symbols through fast Fourier transform. Combining the water-filling algorithm to calculate the entropy corresponding to each subcarrier, using the in-phase and quadrature components of the quadrature amplitude modulation symbol and the average SNR of the received signal as signal features, and using the entropy as the feature label to form a synthetic signal.

4. The adaptive entropy loading method for an optical communication network based on transfer learning according to claim 3, wherein Using the synthetic signal to pre-train the entropy loading model to optimize the model parameters includes: Inputting the signal features in the synthetic signal into the entropy loading model to predict the entropy approaching the channel capacity, using the smooth-L1 function between the entropy prediction value and the feature label as the loss function, and calculating the gradient and backpropagating to update the network parameters of the entropy prediction module in the entropy loading model to achieve parameter optimization of the model.

5. The adaptive entropy loading method for an optical communication network based on transfer learning according to claim 1, wherein Collecting experimental signals through optical communication network experiments includes: Simulating the dynamic changes of the optical communication network through experiments of transmitting discrete multi-tone modulation signals with different frequencies under different link conditions, and after sequentially synchronizing, channel equalizing, and performing fast Fourier transform on the experimental received signal, using the in-phase and quadrature components of the quadrature amplitude modulation symbol and the average SNR of the experimental received signal as signal features, and using the entropy approaching the capacity as the feature label to form an experimental signal.

6. The adaptive entropy loading method for an optical communication network based on transfer learning according to claim 5, wherein Using the experimental signal to perform transfer learning on the pre-trained entropy loading model to achieve parameter fine-tuning includes: Freezing the network layer parameters of the entropy prediction module in the pre-trained entropy loading model, and adding a part of trainable network layers. On this basis, using the experimental signal to fine-tune the newly added network layers to achieve transfer learning.

7. The adaptive entropy loading method for an optical communication network based on transfer learning according to claim 1, wherein Using the entropy loading model with fine-tuned parameters to predict the entropy of the optical communication network signal with unknown channel response includes: Input the signal characteristics of an optical communication network signal with an unknown channel response into the entropy prediction module of the entropy loading model to predict the entropy of the signal.

8. An adaptive entropy loading device for an optical communication network based on transfer learning, characterized in that, It includes: An entropy loading model building unit, which is used to build an entropy loading model including an entropy prediction module and a probability shaping module. Among them, the entropy prediction module is used to predict the entropy approaching the channel capacity based on the signal characteristics, and the probability shaping module is used to perform entropy allocation for each subcarrier using probability amplitude shaping technology to complete the entropy loading of the signal; A synthesized signal generation and model pre-training unit, which is used to generate a synthesized signal by simulating channel characteristics and pre-train the entropy loading model using the synthesized signal to optimize the model parameters; An experimental signal generation and model transfer learning unit: which is used to collect experimental signals through optical communication network experiments and perform transfer learning on the pre-trained entropy loading model using the experimental signals to achieve parameter fine-tuning; An adaptive entropy loading unit, which uses the entropy loading model with fine-tuned parameters to predict the entropy of a signal with an unknown channel response, and performs entropy allocation for each subcarrier using probability amplitude shaping technology to complete the entropy loading of the signal.

9. A computing device, comprising a memory and one or more processors, wherein executable code is stored in the memory, characterized in that, When the one or more processors execute the executable code, it is used to implement the adaptive entropy loading method for an optical communication network based on transfer learning according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Stored thereon is a program, which when executed by a processor, implements the adaptive entropy loading method for an optical communication network based on transfer learning according to any one of claims 1-7.

Images