Joint monitoring method and device for SNRL and SNRNL of WDM system

Abstract

The invention provides a joint monitoring method and device for SNRL and SNRNL of a WDM system, and the method comprises the steps: receiving a signal of a target channel through a coherent receiver in the WDW system, and carrying out the preprocessing of the received signal, and obtaining a constellation point density matrix; respectively training by using the constellation point density matrix to obtain a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation; according to the constellation point density matrix and a teacher model, training to obtain a lightweight multi-task student model after knowledge distillation; and inputting the constellation point density matrix into a multi-task teacher model, a multi-task student model before knowledge distillation and a multi-task student model after knowledge distillation, and respectively carrying out SNRL and SNRNL joint monitoring processing. The teacher model is utilized to guide the training of the student model, and the student model after knowledge distillation remarkably improves the SNRL and SNRNL joint monitoring performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of long-distance optical fiber communication, and in particular relates to a joint monitoring method and device for SNRL and SNRNL of a WDM system. Background Art

[0002] With the rapid growth of Internet traffic, optical communication systems are facing unprecedented challenges. In order to meet the needs of big data, the development of high-speed and large-capacity optical fiber communication is an important way to solve the above challenges. Wavelength Division Multiplexing (WDM) technology, as one of the solutions to alleviate traffic constraints, has greatly expanded system capacity and improved spectrum efficiency. However, this also makes the signal more susceptible to nonlinear effects. Nonlinear noise (NLN) caused by Kerr nonlinear effect has become one of the main obstacles for WDM systems to improve system capacity and transmission distance. In addition, stimulated spontaneous emission (ASE) noise caused by erbium-doped fiber amplifier (EDFA) is usually regarded as linear noise, which also degrades signal quality. SNRL and SNRNL are usually used to quantify NLN noise and ASE noise, which are important parameters reflecting the signal transmission quality in the system. In this case, in order to continuously monitor the signal quality and system performance of the WDM system, it is very necessary to monitor linear and nonlinear noise at the physical layer so as to transmit information to the network control layer in real time, which is conducive to operators to dynamically adjust the physical layer link using feedback information. In addition, the realization of intelligent joint monitoring of SNRL and SNRNL can not only provide operators with comprehensive system analysis, but also greatly improve the efficiency of optical performance monitoring.

[0003] In recent years, researchers have proposed a variety of linear and nonlinear noise joint monitoring schemes based on neural networks. The effectiveness of such schemes is based on a large amount of training data or complex parameter models. However, in actual systems, obtaining a large amount of data will take a lot of time, and complex parameter models will also increase the complexity of the scheme and occupy a lot of memory. In order to overcome the above difficulties, knowledge distillation technology can compress large models into small models to reduce the complexity of neural networks, and the compressed small models still have performance close to that of large models. Although knowledge distillation has made great progress in the classification task model compression task, using knowledge distillation to solve regression problems still faces challenges, especially the exploration of multi-task regression problems. In knowledge distillation, the loss function is used to measure the difference between the student model and the teacher model, helping the student model to obtain knowledge and guidance from the teacher model, and improving the performance and generalization ability of the model. For regression problems, knowledge distillation has been applied in OSNR monitoring schemes. The basic idea of ​​this scheme is to add a "softened" loss function based on the loss function of the teacher model regression prediction and its own label. The softened loss function is the output of the teacher model and the student model after the softmax function, and then solve the KLD. However, the softmax function in this solution only serves to normalize the output prediction value of the network model, and it is difficult to enrich the information carried by the regression prediction value. Furthermore, in the nonlinear Schrödinger equation modeling task, an additional loss function is also added on the basis of the original loss function. The additional loss function is obtained by calculating the MSE of the outputs of the teacher model and the student model, so that the student model can learn the output of the teacher model. However, the output of the teacher model in this solution is also an independent prediction value, which also does not include probability information. In some cases, relying solely on an additional loss function may mislead the training of the student model.

[0004] The existing technologies use knowledge distillation technology to improve the efficiency of optical performance monitoring and reduce complexity, but they do not involve nonlinear noise monitoring that cannot be ignored, and do not consider multi-task regression. According to the search, the multi-task regression problem of realizing SNRL and SNRNL intelligent joint monitoring in WDM systems using knowledge distillation technology has not been involved. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method and device for joint monitoring of SNRL and SNRNL of a WDM system, which improves the joint monitoring performance of SNRL and SNRNL by using a teacher model and a student model before and after knowledge distillation.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention provides a method for jointly monitoring SNRL and SNRNL of a WDM system, comprising the following steps:

[0008] S1. In the WDW system, a coherent receiver is used to receive a signal of a target channel, and the received signal is preprocessed to obtain a constellation point density matrix;

[0009] S2, using the constellation point density matrix to train a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation;

[0010] S3. According to the constellation point density matrix and the teacher model, a lightweight multi-task student model is trained after knowledge distillation.

[0011] S4. Input the constellation point density matrix into the multi-task teacher model, the multi-task student model before knowledge distillation, and the multi-task student model after knowledge distillation, and perform SNRL and SNRNL joint monitoring processing respectively.

[0012] The beneficial effects of the present invention are: the present invention realizes the joint monitoring of the two indicators SNRL and SNRNL in the WDM system based on the knowledge distillation technology, and efficiently transfers knowledge from the complex teacher network to the lightweight student network technology, thereby improving the performance of the joint monitoring of the multi-task student model without changing the computational cost of the scheme. The present invention utilizes the teacher model and the student model before and after knowledge distillation to improve the joint monitoring performance of SNRL and SNRNL.

[0013] In a second aspect, the present invention provides a joint monitoring device for SNRL and SNRNL of a WDM system, comprising:

[0014] A signal preprocessing module is used to receive the signal of the target channel by using a coherent receiver in the WDW system, and preprocess the received signal to obtain a constellation point density matrix;

[0015] The first model training module is used to use the constellation point density matrix to train a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation;

[0016] The second model training module is used to train a lightweight multi-task student model after knowledge distillation based on the constellation point density matrix and the teacher model;

[0017] The monitoring processing module is used to input the constellation point density matrix into the multi-task teacher model, the multi-task student model before knowledge distillation, and the multi-task student model after knowledge distillation, and perform SNRL and SNRNL joint monitoring processing respectively.

[0018] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement a method for intelligent joint monitoring of SNRL and SNRNL of a WDM system based on knowledge distillation.

[0019] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement a method for intelligent joint monitoring of SNRL and SNRNL of a WDM system based on knowledge distillation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is a flow chart of the method of the present invention.

[0021] Figure 2 The overall architecture diagram of the SNRL and SNRNL intelligent joint monitoring based on knowledge distillation provided by the present invention.

[0022] Figure 3 Block diagram of a 28Gbaud 5-channel WDM simulation system in this embodiment.

[0023] Figure 4 Comparison of the MAE of simulation results of the teacher model and the student model before and after knowledge distillation in this embodiment.

[0024] Figure 5 Schematic diagram of MAE at each independent SNRL reference value before and after knowledge distillation in this embodiment.

[0025] Figure 6 Schematic diagram of a 3-channel 28GBaud PDM-16QAM WDM optical transmission experimental system in this embodiment.

[0026] Figure 7 It is a comparison chart of the MAE of the experimental results of the teacher model in this embodiment, and the student model before and after knowledge distillation.

[0027] Figure 8 Schematic diagram of MAE at each independent SNRL reference value before and after knowledge distillation in this embodiment.

[0028] Fig. 9 It is a schematic diagram of the device structure of the present invention. DETAILED DESCRIPTION

[0029] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0030] Before describing the present invention, the following parameters are described:

[0031] WDM (Wavelength division multiplexing): wavelength division multiplexing;

[0032] SNRL (Linear signal-to-noise ratio): linear signal-to-noise ratio;

[0033] SNRNL (Nonlinear signal-to-noise ratio): nonlinear signal-to-noise ratio;

[0034] KLD (Kullback-Leibler divergence): KL divergence.

[0035] Example 1

[0036] Aiming at the problem of intelligent joint monitoring of SNRL and SNRNL in WDM system, the present invention proposes a joint monitoring method of SNRL and SNRNL in WDM system based on knowledge distillation. The flow chart of the method is as follows: Figure 1 As shown. Figure 2 It can be seen that the loss function of the knowledge distillation technology in this method contains three parts.

[0037] like Figure 1 As shown, the present invention provides a joint monitoring method of SNRL and SNRNL of a WDM system, and its implementation method is as follows:

[0038] S1. In the WDW system, a coherent receiver is used to receive a signal of a target channel, and the received signal is preprocessed to obtain a constellation point density matrix;

[0039] In this embodiment, in a WDM system, a coherent receiver receives a signal of a target channel and preprocesses the signal to obtain a constellation point density matrix, and the implementation method thereof is as follows:

[0040] S101, in the WDW system, using a coherent receiver to receive a signal of a target channel;

[0041] S102, according to the received signal, performing DSP processing on the constellation diagram in sequence, including dispersion compensation, polarization demultiplexing, frequency offset estimation, carrier phase recovery and standardization;

[0042] S103, setting a uniform grid, and using the uniform network to divide the constellation diagram;

[0043] S104. Count the number of constellation points in each uniform grid and construct a two-dimensional constellation point density matrix.

[0044] S2. Use the constellation point density matrix to train a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation. The implementation method is as follows:

[0045] S201, using the constellation point density matrix as the input feature of the teacher model, passing through the shared unit and the subtask unit of the teacher model in turn, and saving the complex multi-task teacher model that realizes the joint monitoring of SNRL and SNRNL;

[0046] S202. Use the constellation point density matrix as the input feature of the student model before knowledge distillation, and sequentially pass it through the shared unit and subtask unit of the student model before knowledge distillation to save the lightweight multi-task student model before knowledge distillation that realizes the joint monitoring of SNRL and SNRNL.

[0047] In this embodiment, the generated constellation point density matrix is ​​used as the input feature of the multi-task teacher model, and passes through the shared unit and subtask unit of the teacher model in turn to save the complex multi-task teacher model that can realize the joint monitoring of SNRL and SNRNL. The generated multi-task teacher model guides the training of the lightweight multi-task student model in the knowledge distillation process. In the present invention, the Adam optimizer with a learning rate of 0.001 is selected to update and optimize the tap coefficients, nonlinear phase shift correction factors, and activation function parameters. Batch normalization function and ReLU function are added between the convolution layer and the pooling layer. The batchnormalization function is used to normalize the data set to prevent gradient disappearance and gradient explosion.

[0048] Among them, the shared unit of the multi-task teacher model includes 4 convolutional layers and 3 pooling layers. The parameters of the 4 convolutional layers are: the number of channels of the convolutional layer is 32, and the dimension of the convolutional kernel is [4,4]; the number of channels of the convolutional layer is 64, and the dimension of the convolutional kernel is [3,3]; the number of channels of the convolutional layer is 128, and the dimension of the convolutional kernel is [2,2]; the number of channels of the convolutional layer is 256, and the dimension of the convolutional kernel is [2,2]. The parameters of each pooling layer are [2,2]; the subtask unit sets two monitoring tasks, SNRL and SNRNL, and both are regarded as regression tasks. Each subtask includes 4 hidden layers, and the number of neurons is 200, 200, 50 and 1 respectively.

[0049] In this embodiment, the generated constellation point density matrix is ​​used as the input feature of the student model before multi-task knowledge distillation, and is sequentially passed through the shared unit and subtask unit of the student model before knowledge distillation to save the trained lightweight multi-task student model before knowledge distillation that can realize SNRL and SNRNL joint monitoring. The student model before multi-task knowledge distillation is used to compare the monitoring performance of the multi-task second student model after knowledge distillation. Similar to the teacher model, the learning rate of the Adam optimizer is also set to 0.001, which is used to update and optimize the tap coefficients, nonlinear phase shift correction factors, and activation function parameters. Batch normalization function and ReLU function are used to prevent gradient disappearance and gradient explosion.

[0050] Among them, the student model shared unit before multi-task knowledge distillation includes 3 convolutional layers and 3 pooling layers. The parameters of the 3 convolutional layers are: the number of channels of the convolutional layer is 2, the dimension of the convolutional kernel is [4,4]; the number of channels of the convolutional layer is 4, the dimension of the convolutional kernel is [3,3], and the number of channels of the convolutional layer is 8, the dimension of the convolutional kernel is [2,2]. The parameters of the 3 pooling layers are [2,2], [2,2], and [5,5] respectively; the two task monitorings of SNRL and SNRNL are also regarded as regression tasks. In addition, each subtask includes 2 hidden layers, and the number of neurons is 50 and 1 respectively.

[0051] S3. According to the constellation point density matrix and the teacher model, a lightweight multi-task student model is trained after knowledge distillation.

[0052] In this embodiment, the generated constellation point density matrix is ​​sent to an untrained multi-task student model to implement knowledge distillation technology. The distillation process needs to use the trained multi-task teacher model. During the training process, the loss function is used to guide parameter optimization, measure the difference between the student model and the teacher model after knowledge distillation, and help the student model after knowledge distillation to obtain knowledge and guidance from the teacher model, thereby improving the performance and generalization ability of the model. On the basis of the existing knowledge distillation method for solving regression tasks, the loss function of the present invention adds the KLD term of the similarity matrix between the teacher model and the student model after knowledge distillation. This term, together with the MSE of the output prediction value and the reference value of the student model after knowledge distillation, and the MSE of the prediction value between the teacher model and the student model after knowledge distillation, constitute the loss function of each subtask.

[0053] In this embodiment, during the knowledge distillation process, the total loss function expression is:

[0054]

[0055] Loss SNRL The function expression is:

[0056] Loss SNRL =0.8·(1-β)·Loss hard_SNRL +0.2·(1-β)·Loss soft_SNRL +β·Loss KLD_SNRL

[0057] Among them, Loss SNRL Represents the Loss function of the SNRNL monitoring subtask, Loss soft_SNRL Represents the MSE and Loss of the SNRL prediction values ​​output by the teacher model and the student model KLD_SNRL KLD represents the similarity matrix between the teacher model and the student model. Epoch is the number of iterations in the training process, and epoch is set to 300.

[0058] Loss SNRL The function component expressions are:

[0059]

[0060] Among them, Loss hard_SNRL Represents the MSE of the SNRL prediction value output by the student model and the reference value. i and Represents the output prediction value and reference value of the SNRL monitoring task.

[0061]

[0062] Among them, Loss soft_SNRL represents the MSE of the SNRL prediction values ​​output by the teacher model and the student model, y T and S are the output of the teacher model and the SNRL prediction value of the student model, respectively.

[0063]

[0064] Among them, Loss pkt is the quantitative value of the similarity between two similarity matrices. It is a new loss function in this invention, which is used to make up for the deficiency that the output value of the regression task is a single value without probability information. In this case, the teacher model can more effectively transfer knowledge to the student model, thereby greatly improving the ability of knowledge distillation technology to solve regression problems. log(·) represents logarithmic operation.

[0065] In this embodiment, and Respectively represent the converted two-dimensional symmetric matrix of the teacher model and the student model, which is called the similarity matrix. In the present invention, the hidden layer with 50 neurons in the fully connected layer is selected as the middle layer of the teacher model and the student model, and the expression of the similarity matrix is:

[0066]

[0067]

[0068] in, and They represent the normalized output vectors of the fully connected layers in the teacher model and the student model respectively, and their expressions are:

[0069]

[0070]

[0071] and The two dimensions are expressed as (batch_size, batch_size), batch_size=64.

[0072] Meanwhile, Loss SNRNL The function expression is:

[0073] Loss SNRNL =0.8·(1-β)·Loss hard_SNRL +0.2·(1-β)·Loss soft_SNRL +β·Loss KLD_SNRNL

[0074]

[0075] S4. Input the constellation point density matrix into the multi-task teacher model, the multi-task student model before knowledge distillation, and the multi-task student model after knowledge distillation, and perform SNRL and SNRNL joint monitoring processing respectively.

[0076] In this embodiment, the constellation point density matrix contains both SNRL and SNRNL features, and a multi-task neural network is used to extract the two features to realize joint monitoring of the two tasks. (The multi-task neural network finally has two output layers, corresponding to the two tasks respectively).

[0077] In this embodiment, the block diagram of the 5-channel 28Gbaud WDM simulation system is as follows: Figure 3 As shown. At the transmitting end, the multiplexer multiplexes the optical signals of five different central wavelengths into one signal. The multiplexed signal is amplified by the erbium-doped fiber amplifier and then enters the optical fiber loop. The optical fiber loop consists of 100km of standard single-mode optical fiber and erbium-doped fiber amplifier. The loss, dispersion and nonlinear coefficient of the standard single-mode optical fiber are 0.19dB / km, 16.7ps / (nm·km) and 0.2ps / (km) respectively. 1 / 2and 1.27 / (W·km), and the erbium-doped fiber amplifier is used to compensate for the fiber loss. After transmission through the fiber loop, the optical demultiplexer is responsible for demultiplexing the received signal of each WDM channel at the receiving end. Without loss of generality, the center channel is selected as the target channel and subsequent processing is performed. First, the signal undergoes dispersion compensation, downsampling, and orthogonalization; then, the constant mode algorithm and the multimode algorithm will be used to compensate for polarization mode dispersion and polarization demultiplexing; finally, the frequency offset estimation and carrier phase recovery algorithms are used to compensate for the laser frequency offset and phase noise, respectively. The transmitter pulse shaping filter uses an SRRC filter with a roll-off factor of 0.1, and the noise figure of the erbium-doped fiber amplifier is set to 5dB. In order to ensure that the linear noise and nonlinear noise in the link change simultaneously, the number of fiber reels changes from 4 to 15, with an interval of 1 reel.

[0078] In this embodiment, Figure 4 The MAE comparison of SNRL and SNRNL monitoring in the simulation of this scheme is intuitively demonstrated. Here, the MAE of the teacher model, the student model before and after knowledge distillation are selected to verify the effectiveness of the proposed knowledge distillation scheme. The MAE of the student model and the teacher model before knowledge distillation are both the models trained in step S2, and the MAE of the student model after knowledge distillation is obtained by step S3. Figure 4 (a) describes the simulation results of SNRL. The MAE of the teacher model, the student model before and after knowledge distillation are 0.23dB, 0.38dB, and 0.24dB, respectively. Compared with the student model before knowledge distillation, the MAE of SNRL monitoring after knowledge distillation is reduced from 0.38dB to 0.24dB, and the monitoring performance is improved by 0.14dB. Figure 4 (b) It can be seen that the MAE of the teacher model, the student model before and after distillation in the simulation results of SNRNL are 0.17dB, 0.37dB and 0.22dB respectively. After knowledge distillation, the MAE of the student model SNRNL monitoring dropped from 0.37dB to 0.22dB, and the MAE of the student model SNRNL monitoring dropped by 0.15dB. Figure 4 (a) and (b) show that in the scenario of joint monitoring of SNRL and SNRNL, the student model after knowledge distillation is better than the student model before knowledge distillation and is close to the monitoring results of the teacher model.

[0079] In this embodiment, Figure 5 describes in more detail how knowledge distillation improves the monitoring performance of the student model. Figure 5(a) and (b) describe the changes in MAE before and after knowledge distillation at the reference values ​​of SNRL and SNRNL, respectively. For each independent SNRL value and SNRNL value, the estimated average of the student model after knowledge distillation is smaller than that of the student model before distillation, and the maximum improvement in the monitoring performance of the two parameters of SNRL and SNRNL is 0.38dB and 0.36dB, respectively. In addition, after knowledge distillation, the fluctuation of the MAE curves of the student model SNRL monitoring and SNRNL monitoring is reduced, which also indicates that the stability of the monitoring performance is improved. In addition, in addition to the MAE of all monitoring values ​​at each reference value, the average value of all monitoring values ​​at each reference value can better reflect the degree to which the joint monitoring value of SNRL and SNRNL deviates from the reference value. Figure 5 (c) and (d) show the estimated average values ​​of SNRL and SNRNL of the student model after knowledge distillation. The red lower triangle represents the estimated average values ​​of SNRL and SNRNL, while the blue square represents the reference value of SNRL and SNRNL. Figure 5 (c) and (d) show that the estimated average values ​​of SNRL and SNRNL have the same trend as the reference value. The average deviation errors of the red lower triangle and blue square in (c) and (d) are 0.19dB and 0.12dB respectively. Based on the above simulation results, it can be seen that without increasing the parameters of the student model, knowledge distillation can not only make the monitoring performance of the student model close to the teacher model, but also make the performance of the student model more stable.

[0080] In this embodiment, Figure 6 As shown in the figure, the experiment of 3-channel 28GBaud PDM-16QAM WDM optical transmission system is used to verify the effectiveness of the scheme. First, AWG and three lasers with a spacing of 50GHz are used to generate 28Gbaud PDM-16QAM three-channel multiplexed signals. After being amplified by EDFA, the multiplexed signal enters the optical fiber loop composed of standard SSMF, EDFA and optical filter. Next, the multiplexed signal passes through the demultiplexer, and the function of the demultiplexer is to separate the multiplexed signal into single carrier signals. At the receiving end, the central wavelength of the local oscillator laser is 1550.12nm, and the receiver receives the data of the middle channel and enters the subsequent offline digital signal processing unit. In order to be consistent with the simulation settings, the fiber input power of each channel signal in the experimental system is kept at 2dBm, and the optical fiber transmission distance is increased from 200km to 1200km, with a spacing of 200km. The offline digital signal processing unit includes IQ orthogonalization, dispersion compensation, constant modulus algorithm, multi-mode algorithm, frequency offset estimation and carrier phase estimation modules, which respectively compensate for damages such as IQ imbalance, dispersion, polarization mode dispersion, frequency offset and phase noise.

[0081] In this embodiment, Figure 7It can be seen that the MAE of the teacher model, the student model before and after knowledge distillation can intuitively represent the reduction in monitoring error brought by the knowledge distillation scheme to the student model. Figure 7 (a) depicts the experimental results of SNRL. The MAE of the teacher model, the student model before and after knowledge distillation are 0.4dB, 0.7dB, and 0.58dB, respectively. After the knowledge distillation operation, the MAE of the student model SNRL monitoring is reduced from 0.7dB to 0.58dB. In addition, Figure 7 In (b), the MAE of the teacher model, the student model before and after distillation in the SNRNL experimental results are 0.52dB, 0.8dB, and 0.65dB, respectively. By comparison, before and after knowledge distillation, the student model after knowledge distillation is better than the student model before knowledge distillation, and the MAE of SNRNL monitoring of the student model after knowledge distillation decreases by 0.15dB. It can be seen that in the scenario of joint monitoring of SNRL and SNRNL, knowledge distillation improves the monitoring performance of the student model SNRL and SNRNL by 0.12dB and 0.15dB, respectively.

[0082] To further demonstrate the performance improvement brought by knowledge distillation, Figure 8 The experimental results describe the monitoring results of the student models SNRL and SNRNL before and after knowledge distillation. Figure 8 (a) and (b) respectively describe the changes in MAE of the two models before and after knowledge distillation at the SNRL and SNRNL reference values. For both SNRL and SNRNL monitoring, the estimated average values ​​of the student model after distillation are smaller than the estimated average values ​​before knowledge distillation. Moreover, at each independent SNRL and SNRNL value, the maximum improvement of knowledge distillation on the student model is 0.32dB and 0.23dB, respectively. Figure 8 (c) and (d) depict the estimated average values ​​of the student models SNRL and SNRNL after knowledge distillation in the experimental phase, and Figure 8 (c) and (d) The triangles represent the estimated mean values ​​of SNRL and SNRNL, while the squares represent the reference values ​​of SNRL and SNRNL. Figure 8 (c) and (d) show that the estimated average values ​​of SNRL and SNRNL fluctuate around the reference value. The average deviation errors between the estimated average values ​​of SNRL and SNRNL and the reference value in (c) and (d) are 0.27 dB and 0.36 dB, respectively. Figure 7 and 8 The experimental results show that the knowledge distillation operation can transfer the information of the teacher model to the student model, thereby improving the performance of the student model without changing the parameters of the student model.

[0083] In summary, the present invention proposes an intelligent joint monitoring scheme for SNRL and SNRNL based on knowledge distillation. The knowledge distillation technology is used to realize the transfer of knowledge from the teacher model to the student model, so as to achieve the purpose of improving the performance of the student model. According to the simulation results, after knowledge distillation, the MAE of the student model SNRL and SNRNL monitoring is reduced by 0.14dB and 0.15dB. In addition, in the experimental verification stage, the proposed knowledge distillation scheme brought 0.12dB and 0.15dB performance improvements to the student model SNRL and SNRNL monitoring, respectively. Comprehensive simulation and experimental results show that the knowledge distillation technology proposed in the present invention can be used as an effective model compression scheme to solve the multi-task regression problem.

[0084] Example 2

[0085] like Fig. 9 As shown, the present invention provides a joint monitoring device for SNRL and SNRNL of a WDM system, comprising:

[0086] A signal preprocessing module is used to receive the signal of the target channel by using a coherent receiver in the WDW system, and preprocess the received signal to obtain a constellation point density matrix;

[0087] The first model training module is used to use the constellation point density matrix to train a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation;

[0088] The second model training module is used to train a lightweight multi-task student model after knowledge distillation based on the constellation point density matrix and the teacher model;

[0089] The monitoring processing module is used to input the constellation point density matrix into the multi-task teacher model, the multi-task student model before knowledge distillation, and the multi-task student model after knowledge distillation, and perform SNRL and SNRNL joint monitoring processing respectively.

[0090] like Fig. 9 The joint monitoring device shown can execute the technical solution shown in the joint monitoring method of the above method embodiment. Its implementation principle and beneficial effects are similar and will not be repeated here.

[0091] In an embodiment of the present invention, the present application may divide the functional units according to the joint monitoring method. For example, each function may be divided into each functional unit, or two or more functions may be integrated into one processing unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of a software functional unit. It should be noted that the division of units in the present invention is schematic and is only a logical division. There may be other division methods in actual implementation.

[0092] In this embodiment, in order to realize the principles and beneficial effects of the joint monitoring method, the joint monitoring device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the schematic units and algorithm steps described in the embodiments disclosed in the present invention, the present invention can be implemented in the form of hardware and / or a combination of hardware and computer software. Whether a function is executed in a hardware or computer software driven manner depends on the specific application and design constraints of the technical solution. Different methods can be used for each specific application to implement the described function, but such implementation should not be considered to exceed the scope of this application.

[0093] In this embodiment, the present invention realizes the joint monitoring of the two indicators SNRL and SNRNL in the WDM system based on the knowledge distillation technology, and efficiently transfers knowledge from the complex teacher network to the lightweight student network, thereby improving the performance of the joint monitoring of the multi-task student model without changing the computational cost of the scheme. The present invention utilizes the teacher model and the student model before and after knowledge distillation to improve the joint monitoring performance of SNRL and SNRNL.

[0094] Example 3

[0095] The present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the WDM system SNRL and SNRNL intelligent joint monitoring method based on knowledge distillation as described in any one of Embodiment 1.

[0096] In this embodiment, the electronic device may include: a processor, a memory, a bus and a communication interface. The processor, the communication interface and the memory are connected via a bus. The memory stores a computer program that can be run on the processor. When the processor runs the computer program, it executes part or all of the steps of the WDM system SNRL and SNRNL intelligent joint monitoring method based on knowledge distillation provided in the aforementioned embodiment 1 of the present application.

[0097] Example 4

[0098] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement any of the knowledge distillation-based WDM system SNRL and SNRNL intelligent joint monitoring methods described in Example 1.

[0099] The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk, and the readable storage medium can be any available medium that can be accessed by a general or special-purpose computer. The readable storage medium is coupled to the processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. The readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application-specific integrated circuit (ASIC). The processor and the readable storage medium can also exist as discrete components in the point de-registration system.

[0100] The embodiments of the present invention may provide methods, devices, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. With reference to the process and / or block diagram of the method, device (device), and computer program product according to the embodiments of the present invention, it should be understood that each process and / or block diagram in the flow chart and / or block diagram, as well as the combination of the flow chart and / or block diagram, can be implemented by computer program instructions, and these computer program instructions can be provided to a computer-readable memory of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device that works in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which is implemented in the flow chart. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing one or more processes and / or blocks in the flowchart. Figure 1 A step that specifies a function in one or more boxes.

[0101] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A joint monitoring method for SNRL and SNRNL of a WDM system, characterized in that: The following steps are involved: S1. In the WDW system, a coherent receiver is used to receive a signal of a target channel, and the received signal is preprocessed to obtain a constellation point density matrix; S2, using the constellation point density matrix to train a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation; S3. According to the constellation point density matrix and the teacher model, a lightweight multi-task student model is trained after knowledge distillation. S4. Input the constellation point density matrix into the multi-task teacher model, the multi-task student model before knowledge distillation, and the multi-task student model after knowledge distillation, and perform SNRL and SNRNL joint monitoring processing respectively.

2. The joint monitoring method of SNRL and SNRNL of a WDM system according to claim 1, characterized in that: The step S1 comprises the following steps: S101, in the WDW system, using a coherent receiver to receive a signal of a target channel; S102, according to the received signal, performing DSP processing on the constellation diagram in sequence, including dispersion compensation, polarization demultiplexing, frequency offset estimation, carrier phase recovery and standardization; S103, setting a uniform grid, and using the uniform network to divide the constellation diagram; S104. Count the number of constellation points in each uniform grid and construct a two-dimensional constellation point density matrix.

3. The joint monitoring method of SNRL and SNRNL of a WDM system according to claim 1, characterized in that: The step S2 comprises the following steps: S201, using the constellation point density matrix as the input feature of the teacher model, passing through the shared unit and the subtask unit of the teacher model in turn, and saving the complex multi-task teacher model that realizes the joint monitoring of SNRL and SNRNL; S202. Use the constellation point density matrix as the input feature of the student model before knowledge distillation, and sequentially pass through the shared unit and subtask unit of the student model before knowledge distillation to save the lightweight multi-task student model before knowledge distillation that realizes the joint monitoring of SNRL and SNRNL.

4. The joint monitoring method of SNRL and SNRNL of a WDM system according to claim 3, characterized in that: The shared unit of the teacher model includes 4 convolutional layers and 3 pooling layers, and the parameters of the 4 convolutional layers are: the number of channels of the convolutional layer is 32, and the dimension of the convolutional kernel is [4,4]; the number of channels of the convolutional layer is 64, and the dimension of the convolutional kernel is [3,3]; the number of channels of the convolutional layer is 128, and the dimension of the convolutional kernel is [2,2]; the number of channels of the convolutional layer is 256, and the dimension of the convolutional kernel is [2,2]; The subtask unit of the teacher model is provided with two monitoring tasks, SNRL and SNRNL. Each subtask unit includes three hidden layers, and the number of neurons in the four hidden layers is 200, 200, 50 and 1 respectively.

5. The joint monitoring method of SNRL and SNRNL of a WDM system according to claim 3, characterized in that: The shared unit of the student model before the knowledge distillation includes 3 convolutional layers and 3 pooling layers, and the parameters of the 3 convolutional layers are: the number of channels of the convolutional layer is 2, and the dimension of the convolutional kernel is [4, 4]; the number of channels of the convolutional layer is 4, and the dimension of the convolutional kernel is [3, 3]; the number of channels of the convolutional layer is 8, and the dimension of the convolutional kernel is [2, 2], and the parameters of the 3 pooling layers are [2, 2], [2, 2], and [5, 5] respectively; The subtask unit of the student model before the knowledge distillation is set with two tasks, SNRL and SNRNL. Each of the subtask units includes two hidden layers, and the number of neurons in the two hidden layers is 50 and 1 respectively.

6. The joint monitoring method of SNRL and SNRNL of a WDM system according to claim 1, characterized in that: The lightweight multi-task student model obtained after knowledge distillation in step S3 has a loss function in the training process as follows: Loss SNRL =0.8·(1-β)·Loss hard_SNRL +0.2·(1-β)·Loss soft_SNRL +β·Loss KLD_SNRL Loss SNRNL =0.8·(1-β)·Loss hard_SNRL +0.2·(1-β)·Loss soft_SNRL +β·Loss KLD_SNRNL Among them, Loss total Represents the loss function of multi-task student model training in the knowledge distillation process, Loss SNRL represents the loss function of the SNRL monitoring subtask, Loss SNRNL represents the loss function of the SNRNL monitoring subtask, β represents the coefficient before the loss function, and Loss hard_SNRL Represents the MSE and Loss of the SNRL prediction value output by the student model and the reference value soft_SNRL Represents the MSE and Loss of the SNRL prediction values ​​output by the teacher model and the student model KLD_SNRL KLD, Loss represents the similarity matrix between the teacher model and the student model in the SNRL monitoring subtask KLD_SNRNL represents the KLD of the similarity matrix between the teacher model and the student model in the SNRNL monitoring subtask, i represents the i-th output prediction value, n represents the total number of output prediction values, and y i and They represent the output prediction value and reference value of the SNRL monitoring subtask, y T and S Represent the output of the teacher model and the SNRL prediction value of the student model, respectively. B represents the total number of similarity matrices. j represents the jth similarity matrix. and represents the two-dimensional symmetric matrix of the converted teacher model and student model, called the similarity matrix, log(·) represents the logarithmic operation, and Represent the normalized output vectors of the fully connected layers in the teacher model and the student model, respectively. and Respectively and The transpose of , where N is the total number of elements in the vector.

7. A joint monitoring device for the joint monitoring method of SNRL and SNRNL of a WDM system according to any one of claims 1 to 6, characterized in that: include: A signal preprocessing module is used to receive the signal of the target channel by using a coherent receiver in the WDW system, and preprocess the received signal to obtain a constellation point density matrix; The first model training module is used to use the constellation point density matrix to train a complex multi-task teacher model and a lightweight multi-task student model before knowledge distillation; The second model training module is used to train a lightweight multi-task student model after knowledge distillation based on the constellation point density matrix and the teacher model; The monitoring processing module is used to input the constellation point density matrix into the multi-task teacher model, the multi-task student model before knowledge distillation, and the multi-task student model after knowledge distillation, and perform SNRL and SNRNL joint monitoring processing respectively.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor executes the program to implement the WDM system SNRL and SNRNL intelligent joint monitoring method based on knowledge distillation as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program is executed by a processor to implement the WDM system SNRL and SNRNL intelligent joint monitoring method based on knowledge distillation as described in any one of claims 1-6.

Images