Direct modulation laser modeling method based on deep learning

Abstract

The invention discloses a direct modulation laser modeling method based on deep learning. The method comprises the following steps: constructing a direct modulation laser time sequence data set and dividing the direct modulation laser time sequence data set into a training data set and a test data set according to a preset proportion; constructing a deep learning model based on the bidirectional long and short time memory network structure; training the deep learning model based on the training data set, optimizing the training loss by adopting an adaptive moment estimation method until the accuracy requirement is met, and evaluating by adopting the test data set; and modeling the directly modulated laser based on the trained deep learning model. According to the model, modeling can be successfully carried out on nonlinear influence and noise interference introduced by the direct modulation laser, the generated output waveform and frequency spectrum have high accuracy, meanwhile, the real-time requirement for modeling of the direct modulation laser is met, and the model can be used for optical communication system design and dynamic performance optimization and has wide application prospects.

Description

Technical Field

[0001] The present invention belongs to the field of optical communication technology, and in particular relates to a direct modulation laser modeling method based on deep learning. Background Art

[0002] Directly modulated lasers (DMLs) are key components in modern optical communication systems, featuring compact structure and low cost. At present, the modeling methods for directly modulated lasers are mainly based on traditional numerical simulation methods. Such methods describe the dynamic characteristics of carrier density and photon density inside directly modulated lasers by solving the rate equations of lasers through numerical simulation. Although this method can simulate the dynamic behavior of lasers more accurately, its computational complexity is high, and the adjustment and optimization of parameters requires strong mathematical and physical knowledge support, which makes it difficult to meet the needs of real-time applications. In addition, this method has poor adaptability to environmental factors such as external noise and power supply voltage fluctuations, which limits its application in complex practical scenarios.

[0003] To address this shortcoming, a deep learning method was proposed for modeling directly modulated lasers. This method takes advantage of the ability of deep learning to extract features from large amounts of data. In the prior art, a Transformer-based network was used to fit the dynamic characteristics of directly modulated lasers. Its advantage is that it can automatically learn the dynamic response characteristics of directly modulated lasers in a data-driven manner without explicitly defining complex physical models. However, these early studies still have some limitations, such as the model's inaccurate modeling of nonlinear characteristics under high baud rate conditions, or the model requires a large data set, which is computationally complex and difficult to meet the needs of real-time response. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a direct modulated laser modeling method based on deep learning. Through a lightweight improved BiLSTM network, the modeling of the direct modulated laser is completed. It can maintain high-precision prediction while ensuring real-time response, thereby solving the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides a direct modulated laser modeling method based on deep learning, comprising:

[0006] Construct a directly modulated laser timing data set and divide it into a training data set and a test data set according to a preset ratio;

[0007] Constructing a deep learning model based on a bidirectional long short-term memory network structure; training the deep learning model based on a training data set, optimizing the training loss using an adaptive moment estimation method until accuracy requirements are met and evaluating using a test data set;

[0008] Modeling of directly modulated lasers based on the trained deep learning model.

[0009] Optionally, the input current data and output optical power data of the directly modulated laser are collected and preprocessed by simulating a 4th-order pulse amplitude modulation signal, and a timing data set of the directly modulated laser is constructed based on the preprocessed data, wherein the input current data and the output optical power data are timing data.

[0010] Optionally, after the input current data and the output optical power data are normalized, they are divided into windows according to a preset window length.

[0011] Optionally, the deep learning model includes a bidirectional long short-term memory network structure, a feature recalibration mechanism structure and a nonlinear regression module.

[0012] Optionally, the feature recalibration mechanism structure is expressed as follows:

[0013] s t =σ(W expand ·δ(W compress ·x t ))

[0014]

[0015] Among them, s t is the importance score of the feature, x t is the input data, W expand and W compress They are the expansion weight matrix and the compression weight matrix, δ is the ReLU activation function, and x′ is the adjusted feature.

[0016] Optionally, the process of training the deep learning model based on the training data set includes:

[0017] The data in the training data set is organized according to a preset batch size, and the deep learning model is trained through the organized data; the hidden layer size of the deep learning model is 128, the number of hidden layers is 2, and the window length is 1024.

[0018] The present invention also discloses a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned direct modulated laser modeling method based on deep learning.

[0019] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned direct modulation laser modeling method based on deep learning are implemented.

[0020] The present invention also discloses a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned direct modulated laser modeling method based on deep learning.

[0021] Compared with the prior art, the present invention has the following advantages and technical effects:

[0022] The present invention adopts a deep learning method, improves the BiLSTM network, combines the feature recalibration mechanism and nonlinear regression technology, and realizes high-precision, low-computational complexity DML dynamic modeling. Compared with traditional numerical methods and other deep learning models (such as standard BiLSTM and recurrent neural network RNN), the present invention reduces the computational complexity while ensuring the prediction accuracy. Experimental results show that the model of the present invention can successfully model the nonlinear effects and noise interference introduced by the direct modulated laser, and the generated output waveform and spectrum have high accuracy. At the same time, it meets the real-time requirements of DML modeling, and can be used for optical communication system design and dynamic performance optimization, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0024] Figure 1 A schematic diagram of a directly modulated laser modeling system according to an embodiment of the present invention;

[0025] Figure 2 This is a BiLSTM network structure diagram of an embodiment of the present invention;

[0026] Figure 3 A structural diagram of a feature recalibration mechanism according to an embodiment of the present invention;

[0027] Figure 4 is a model structure diagram of an embodiment of the present invention;

[0028] Figure 5 The test output waveform diagram of the model before and after improvement of the embodiment of the present invention;

[0029] Figure 6 This is a test output spectrum diagram before and after the model improvement of the embodiment of the present invention. DETAILED DESCRIPTION

[0030] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0032] Embodiment 1

[0033] like Figure 1-6 As shown, this embodiment provides a direct modulation laser modeling method based on deep learning, including:

[0034] Construct a directly modulated laser timing data set and divide it into a training data set and a test data set according to a preset ratio;

[0035] Specifically, the input current data and output optical power data of the direct modulated laser are collected and preprocessed by simulating a 4th-order pulse amplitude modulation signal, and a direct modulated laser timing data set is constructed according to the preprocessed data, wherein the input current data and the output optical power data are timing data. The preprocessing process includes: after normalizing the input current data and the output optical power data, window division is performed according to a preset window length.

[0036] For example, the required data is recorded according to the sampling period t. Usually, a smaller period is set to obtain more detailed data. In this embodiment, 1.25×10 -11 s, save the data in chronological order as a sequence, and then normalize the data to prevent the data from deviating too far from the average value and causing the model to converge slowly. The normalization calculation formula is:

[0037]

[0038] In the formula, x′ represents the normalized data, and x represents the data before normalization. After the data is normalized, the window is divided, and the window length is set to L. Further, the read batch size B is set, and each B window is combined into a matrix of shape B×L×S, where S is the number of features in the data. Furthermore, the final data is randomly divided into a training data set and a test data set in a ratio of 8:2.

[0039] Build a deep learning model based on a bidirectional long short-term memory network structure;

[0040] Specifically, the deep learning model includes a bidirectional long short-term memory network structure, a feature recalibration mechanism structure and a nonlinear regression module, such as Figure 4 shown.

[0041] Furthermore, according to the studied direct modulated laser modeling system ( Figure 1), build the proposed deep learning model, where the backbone network BiLSTM structure is as follows Figure 2 As shown, it includes a forget gate f t , input gate i t and output gate o t These three gate structures. The forget gate determines which information should be discarded from the memory, the input gate determines which new information will be stored in the memory from the candidate memory, and the output gate determines the output of the final hidden state through the memory. Its calculation formula includes the following parts:

[0042] f t =σ(W f ·[h t-1 , x t ]+b f )

[0043] i t =σ(W i ·[h t-1 , x t ]+b i )

[0044]

[0045] o t =σ(W o ·[h t-1 , x t ]+b o )

[0046] h t =o t tanh(C t )

[0047] Where W f , W i , W C , W o are the weight matrices of the forget gate, input gate, candidate memory, and output gate, respectively, and b f 、b i 、b C 、b o are the biases of the forget gate, input gate, candidate memory, and output gate, respectively. t and are memory and candidate memory respectively, σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function. The structure of the feature recalibration mechanism is as follows Figure 3 As shown, the calculation process is as follows:

[0048] s t =σ(W expand ·δ(W compress ·x t))

[0049] where s t is the importance score of the feature, x t is the input data, δ is the ReLU activation function, W expand and W compress They are the expansion weight matrix and the compression weight matrix. The initial values ​​of the above two matrices are random values. During the training process of the deep learning model, the values ​​of the weight matrices are continuously iteratively updated until the optimal solution is reached.

[0050] The features are adjusted using the above importance scores as follows:

[0051]

[0052] Where x' is the adjusted feature, and the nonlinear regression module that can be further learned has its theoretical basis based on the Kolmogorov-Arnold representation theorem, which is equivalent to implementing the above theorem through the nonlinear regression module of deep learning.

[0053] The available formula is as follows:

[0054]

[0055] In the formula, φ q and φ q,p It is an abstract function that includes different calculations. The function of the above abstract function is to map the input to another value range. The mapping of the value range is expressed as follows:

[0056] According to the mapping of the value range, the above formula can be described as: q,p Change x, which was originally in the interval [0,1] p Mapped to the real number field R, φ q Then pass through φ q,p The output of is taken as input and mapped to R as well. After multiple summations, an abstract function f is finally formed. The function of f is to map the n-dimensional input x with a value range of [0,1] to R.

[0057] Training the deep learning model based on the training data set, optimizing the training loss using an adaptive moment estimation method until the accuracy requirement is met and evaluating using a test data set;

[0058] Modeling of directly modulated lasers based on the trained deep learning model.

[0059] Specifically, the process of training the deep learning model based on the training data set includes:

[0060] The data in the training data set is organized according to a preset batch size, and the deep learning model is trained by the organized data.

[0061] Furthermore, after completing the model construction, the model hyperparameters are set, where the hidden layer size is 128, the number of hidden layers is 2, and the window length is 1024. The proposed model is trained on the PAM4 training set, with 100 training iterations, a batch size of 32, an initial learning rate of 0.001, and the adaptive moment estimation method (Adam) is used to optimize the training loss, which is defined as:

[0062]

[0063] In the formula is the predicted value, y i is the true value. After the training is completed, the trained model is tested using the test data set, including observing the consistency of the time domain waveform and the consistency of the spectrum to verify whether accurate direct modulated laser modeling can be achieved. If not, the model is initialized and retrained to finally obtain a direct modulated laser model that meets the requirements.

[0064] After training, this embodiment evaluates the proposed direct modulated laser modeling method on the PAM4 test set. For fair comparison, this embodiment tests the BiLSTM model before improvement and the proposed improved BiLSTM model. By comparing the predicted waveform and spectrum with the labeled waveform and spectrum, the model's ability to model the direct modulated laser is observed. The waveform and spectrum of the test results are visualized as follows: Figure 5 , Figure 6 shown. Figure 5 The results show that the unimproved BiLSTM can capture the waveform of the PAM4 signal in terms of overall trend, but there are significant errors in details, especially the poor performance of the noise and nonlinear effects introduced by the directly modulated laser. The improved BiLSTM model can fit the actual waveform of the PAM4 signal more accurately. Figure 6 The spectrograms of the waveforms predicted by the BiLSTM model before and after the improvement are shown. The test results show that the improved BiLSTM model exhibits excellent fitting ability in the frequency range of 0-10GHz and can accurately capture the main frequency characteristics of the PAM4 signal. In addition, in the high-frequency range (10-20GHz), the improved BiLSTM model still maintains a low error level and successfully simulates the noise influence introduced by the laser. This shows that the model has strong robustness in the complexity management of high-frequency signals, while ensuring the fitting accuracy in the low-frequency range. In contrast, the performance of the unimproved BiLSTM model in the high-frequency part is significantly reduced, and it fails to effectively capture the high-frequency nonlinear characteristics and noise interference in the laser signal.

[0065] This embodiment also discloses a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned direct modulated laser modeling method based on deep learning.

[0066] This embodiment also discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned direct modulated laser modeling method based on deep learning are implemented.

[0067] This embodiment also discloses a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned direct modulated laser modeling method based on deep learning.

[0068] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A direct modulated laser modeling method based on deep learning, characterized in that: The following steps are involved: Construct a directly modulated laser timing data set and divide it into a training data set and a test data set according to a preset ratio; Build a deep learning model based on a bidirectional long short-term memory network structure; Training the deep learning model based on the training data set, optimizing the training loss using an adaptive moment estimation method until the accuracy requirement is met and evaluating using a test data set; Modeling of directly modulated lasers based on the trained deep learning model.

2. The direct modulated laser modeling method based on deep learning according to claim 1, characterized in that: The input current data and output optical power data of the directly modulated laser are collected by simulating a 4th-order pulse amplitude modulation signal and preprocessed, and a timing data set of the directly modulated laser is constructed according to the preprocessed data, wherein the input current data and the output optical power data are timing data.

3. The direct modulated laser modeling method based on deep learning according to claim 1, characterized in that: After the input current data and the output optical power data are normalized, they are divided into windows according to the preset window length.

4. The direct modulated laser modeling method based on deep learning according to claim 1, characterized in that: The deep learning model includes a bidirectional long short-term memory network structure, a feature recalibration mechanism structure and a nonlinear regression module.

5. The direct modulated laser modeling method based on deep learning according to claim 4, characterized in that: The feature recalibration mechanism structure is shown as follows: s t =σ(W expand ·δ(W compress ·x t )) Among them, s t is the importance score of the feature, x t is the input data, W expand and W compress They are the expansion weight matrix and the compression weight matrix, δ is the ReLU activation function, and x' is the adjusted feature.

6. The direct modulated laser modeling method based on deep learning according to claim 1, characterized in that: The process of training the deep learning model based on the training data set includes: The data in the training data set is organized according to a preset batch size, and the deep learning model is trained through the organized data; the hidden layer size of the deep learning model is 128, the number of hidden layers is 2, and the window length is 1024.

7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the deep learning-based direct modulated laser modeling method described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the deep learning-based direct modulated laser modeling method described in any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the deep learning-based direct modulated laser modeling method described in any one of claims 1 to 6 are implemented.

Images