Method for extracting spectral parameters of DFB laser

Abstract

The invention relates to the field of laser spectrum parameter extraction, and discloses a DFB laser spectrum parameter extraction method comprising the following steps: constructing a deep learning model, and training a corresponding relation between spectrum parameters and spectrum extraction information; testing a DFB laser of which parameters are to be extracted by utilizing a spectrograph, and extracting information by taking a tested spectrum as a target spectrum; inputting the initialized spectrum parameters into a learning model to predict spectrum extraction information, performing error analysis on the spectrum extraction information output by the model and target spectrum extraction information, performing population evolution by using a differential evolution algorithm, and inputting the evolved parameters into the model until target precision or iteration times are reached, and completing the extraction of spectrum parameters. According to the method disclosed by the invention, the parameter extraction time is greatly shortened, and the calculation cost is reduced; spectral parameters can be rapidly extracted, and the situation of local optimum is avoided; and batch extraction of the spectrum parameters of the DFB laser can be realized in a short time.

Description

Technical Field

[0001] The invention relates to the field of laser spectrum parameter extraction, and in particular to a method for extracting DFB laser spectrum parameters. Background Art

[0002] With the advantages of low threshold current, small size and low power consumption, DFB lasers have been widely used in high-speed optical communication systems, sensor systems and medical fields. However, in the actual production process, the material parameters and geometric dimensions of the lasers will deviate from the design values ​​due to the influence of the manufacturing process, manufacturing equipment and even different batches. These deviations will cause the test results of the device to be inconsistent with the simulation results, so it is necessary to extract the actual physical parameters based on performance testing. Parameter extraction is a key method to identify defects in the process, improve laser design and verify theoretical models.

[0003] At present, parameter extraction methods usually combine traditional optimization algorithms with spectral numerical calculation methods. The optimization algorithm is used to iteratively modify and simulate the spectrum-related parameters step by step until the error between the calculated spectrum and the target spectrum meets the requirements or the number of iterations is reached. If the traditional optimization algorithm is run on the condition of iteration, it is very time-consuming and requires high computational costs in order to obtain the interaction of multiple physical states and the intrinsic trends of the device. It is difficult to achieve large-scale parameter extraction based on this method. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for extracting spectrum parameters of a DFB laser, so as to solve the problems that the existing parameter extraction methods are time-consuming and computationally expensive.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A method for extracting spectrum parameters of a DFB laser comprises the following steps: Step 1, constructing a data set using spectral parameters and corresponding spectral extraction information, and constructing a deep learning model, using the spectral parameters as the input of the model, and the spectral extraction information as the output of the model, and using the data set to train and test the model; the spectral parameters include grating period, effective refractive index, coupling coefficient, phase and group refractive index; the spectral extraction information includes the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum; Step 2, using a spectrometer to test the DFB laser whose parameters are to be extracted, obtaining spectral extraction information, and using it as target spectral extraction information; using a differential evolution algorithm to extract the spectral parameters of the DFB laser, firstly initialize the population, and input the initialized spectral parameters into a qualified model to predict the spectral extraction information, and perform error analysis between the spectral extraction information output by the model and the target spectral extraction information, and then drive the population to continuously evolve through the process of mutation, crossover and selection, and input the evolved parameters into the model to predict the spectral extraction information, and then perform error analysis between the spectral extraction information output by the model and the target spectral extraction information, until the target accuracy or number of iterations is reached, and the extraction of spectral parameters is completed.

[0007] In the above scheme, in step 1, when constructing the data set, first set the range of the spectral parameters, then uniformly select points for each parameter based on the set range, and then randomly combine the five spectral parameters to obtain several groups of spectral parameters. For each group of spectral parameters, the spectrum of the DFB laser is calculated using the transfer matrix method; based on the calculated spectrum, the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum are extracted to obtain the corresponding spectrum extraction information.

[0008] In the above scheme, in step 1, when constructing a data set, the spectral parameters are standardized and preprocessed, the spectral extraction information is normalized and preprocessed, the preprocessed spectral parameters and spectral extraction information are used to construct a data set, and the data set is divided into a training set, a validation set, and a test set.

[0009] In the above scheme, in step 1, during the model training process, a loss function is used to quantify the difference between the spectral extraction information predicted by the model and the spectral extraction information in the training set; during the training process, the data in the validation set is used to measure whether the training process is overfitting; after the model training is completed, the model is tested using the data in the test set. If the test is qualified, the next step is entered. If it fails, the model is retrained.

[0010] In the above scheme, in step 1, the constructed deep learning model includes an input layer, 5 hidden layers and an output layer; the input layer contains 5 neurons, corresponding to 5 spectral parameters respectively; the output layer contains 22 neurons, corresponding to the wavelength coordinates and amplitudes of 22 spectral peaks and valleys respectively; the number of neurons in the 5 hidden layers are 100, 200, 300, 200, and 100 respectively; batch normalization is introduced before each hidden layer to adjust the feature distribution and improve convergence stability.

[0011] In a further technical solution, the activation function used by the deep learning model is the ReLu activation function.

[0012] In a further technical solution, during the model training process, the loss functions used are mean square error and mean absolute error, and the formula is as follows: ; ; in, is the mean square error, is the mean absolute error, n is the total number of samples, It is The theoretical value of samples is It is The predicted value of a sample.

[0013] In a further technical solution, the batch normalization method is as follows: In the training batches of dimensions There are A samples in is the first The first of the training dimensions samples; To standardize: ; in, For The data after standardization; is a very small constant used to prevent the denominator from being zero. For this training batch The mean of is the variance of the training batch; ; ; After scaling and translation factors: ; in, for The data after scaling and translation factor transformation is the input of the hidden layer. and denote the parameter vectors for scaling and translation respectively.

[0014] In the above scheme, in step 2, the spectrum extraction information of the target test is the lasing spectrum of the DFB laser near the threshold tested by the spectrometer.

[0015] In the above scheme, in step 2, the differential evolution algorithm includes four steps: initialization, mutation, crossover and selection. The specific process is as follows:

[0016] (1) Initialization For a 5-dimensional search space, population initialization is expressed as: ; in, Is the first in the G generation candidate solutions, each of which contains five spectral parameters, namely a set of values ​​of grating period, effective refractive index, phase, coupling coefficient and group refractive index. N is the size of the population. represents the value of the candidate solution in the jth dimension; Initialize candidate solutions randomly in the search space Initial value of dimension ,in and In the The minimum and maximum values ​​of the search range for the dimension, is a uniform random number in the range of 0 to 1; After the population is initialized, the initialized spectral parameters are input into the deep learning model, and then the mean square error is used to calculate the error between the spectral extraction information of the actual test and the spectral extraction information predicted by the model;

[0017] (2) Variation The mutation operation is used to generate new candidate solutions: ; in, It is The mutation vector corresponding to the candidate solution is , and are three different candidate solutions in the population, and F is a scaling factor used to control the scale of the difference vector;

[0018] (3) Cross The mutated candidate solution and the original candidate solution are subjected to a binomial crossover operation to generate the test individuals, which can be expressed as: ; in, is the crossover probability, is a randomly selected dimension, ensuring that at least one dimension changes; When the evolutionary generation is G, The candidate solution The value of the dimension;

[0019] (4) Selection Finally, the selection operation between the mutated candidate solution and the original candidate solution is performed. The specific formula is as follows: ; in, is the objective function, Indicates the number of evolutionary generations in the parent population when G+1 Candidate solutions, Indicates the number of evolution generations in the crossover population when G A candidate solution, if If the fitness of is better, it will enter the next generation, otherwise the original candidate solution will be retained.

[0020] Through the above technical solution, the method for extracting DFB laser spectrum parameters provided by the present invention has the following beneficial effects:

[0021] The parameter extraction method provided by the present invention can quickly extract spectral parameters when a laser spectrum is available, avoiding the need for a large number of numerical simulation operations when extracting parameters in traditional optimization methods, greatly reducing the time for parameter extraction and reducing the computational cost; compared with traditional optimization algorithms, the fusion of deep learning and differential evolution algorithms can quickly extract spectral parameters and avoid falling into the local optimal situation; batch extraction of DFB laser spectral parameters can be achieved in a relatively short time. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0023] Figure 1 A schematic diagram of a process for extracting spectral parameters of a DFB laser using a differential evolution algorithm disclosed in an embodiment of the present invention; Figure 2 Schematic diagram of the uniform grating DFB laser structure; Figure 3 Schematic diagrams comparing spectrum extraction information predicted based on two sets of parameters in a test set and spectrum extraction information in a test set in an embodiment of the present invention; (a) is a schematic diagram for comparing parameter sample 1, and (b) is a schematic diagram for comparing parameter sample 2; Figure 4 The comparison between the spectrum extraction information predicted based on the test spectrum extraction parameters and the test spectrum extraction information in the embodiment of the present invention; Figure 5 It is a schematic diagram of comparison between a spectrum calculated by using a transfer matrix method based on extraction parameters and a test spectrum in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0025] The present invention provides a method for extracting spectrum parameters of a DFB laser, comprising the following steps: 1. Set the range of spectral parameters, which include grating period, effective refractive index, coupling coefficient, phase and group refractive index; then uniformly select points for each parameter based on the set range, and then randomly combine these five parameters to obtain several groups of spectral parameters, and use the transfer matrix method to calculate the spectrum of the DFB laser for each group of spectral parameters; The specific method of calculating the spectrum of the DFB laser using the transfer matrix method can be found in patent CN 119358427 A, which will not be described in detail here.

[0026] 2. According to the calculated spectrum of the DFB laser, the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum are extracted, and then the spectrum parameters and the corresponding spectrum extraction information are preprocessed respectively; the preprocessed spectrum parameters and spectrum extraction information are used to construct a data set, and the data set is divided into a training set, a validation set, and a test set;

[0027] The method for preprocessing spectral parameters is standardization, and the method for preprocessing spectral extraction information is maximum and minimum value normalization. The specific methods are as follows: Standardizing the input data (spectral parameters) is to adjust each feature dimension to have a mean of 0 and a variance of 1. Assume that Samples , for each dimension feature , first calculate its mean and variance: ; ; The characteristics Subtract the mean and divide by the standard deviation to get the new eigenvalues : ; Normalizing the output data (spectral extraction information) to the maximum and minimum values ​​is to normalize the value range of each feature to Assume that Samples , for each dimension feature The new features obtained after normalization are: ; in, and are the minimum and maximum values ​​in this dimension respectively.

[0028] 3. Build a deep learning model and use the data in the training set to train the model, where the spectral parameters are used as the model input and the spectral extraction information is used as the model output. During the model training process, a loss function is used to quantify the difference between the spectral extraction information predicted by the model and the spectral extraction information in the training set. During the training process, the data in the validation set is used to measure whether the training process is overfitting.

[0029] The constructed deep learning model includes an input layer, 5 hidden layers and an output layer; the input layer contains 5 neurons, corresponding to 5 spectral parameters respectively; the output layer contains 22 neurons, corresponding to the wavelength coordinates and amplitudes of 22 spectral peaks and valleys respectively; the number of neurons in the 5 hidden layers is 100, 200, 300, 200, and 100 respectively; batch normalization is introduced before each hidden layer to adjust the feature distribution and improve the convergence stability. The activation function used in the deep learning model is the ReLu activation function.

[0030] 60,000 samples were used to build the deep learning model. During the model parameter optimization process, the Nadam optimizer was used to adjust the weights and biases of the neurons to minimize the loss function. The learning rate of the optimizer was 0.001. The loss functions used in the optimization process were the mean square error and the mean absolute error. The two jointly quantified the difference between the wavelength coordinates and amplitudes of the predicted spectral peaks and valleys in the model and the training set data. The formula is as follows: ; ; in, is the mean square error, is the mean absolute error, n is the total number of samples, It is The theoretical value of samples is It is The predicted value of a sample.

[0031] The batch normalization method is as follows: In the training batches of dimensions There are A samples in is the first The first of the training dimensions samples; To standardize: ; in, For The data after standardization; is a very small constant used to prevent the denominator from being zero. For this training batch The mean of is the variance of the training batch; ; ; After scaling and translation factors: ; in, for The data after scaling and translation factor transformation is the input of the hidden layer. and denote the parameter vectors for scaling and translation respectively.

[0032] 4. After the model training is completed, the model is tested using the data in the test set. If the test is qualified, proceed to the next step. If it fails, retrain the model.

[0033] The method provided by the present invention is applied to a uniform grating DFB laser, and its structural schematic diagram is shown in FIG. Figure 2 The ranges of spectral parameters given when generating the dataset are shown in Table 1.

[0034] Table 1 Ranges of spectral parameters

[0035] Within the parameter range in Table 1, each parameter is uniformly selected, and then the five parameters are randomly combined to obtain several groups of spectral parameters. Then, 78,000 spectral samples are calculated by the transfer matrix method, and the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum are extracted to construct a data set together with the relevant parameters of the spectrum. A deep learning model is constructed based on 60,000 samples, 9,000 samples are used to verify the model, and 9,000 samples are used to evaluate the generalization ability of the constructed model. Two parameter samples are randomly selected from the test set and input into the constructed deep learning model to predict the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum. The results are shown in Figure 2. Figure 3 As shown in (a) and (b), the spectral information in the test set is the dotted line, and the model's prediction information is the solid line. Figure 3 It can be seen that the spectrum extraction information predicted by the constructed model is highly consistent with the data in the dataset, which intuitively shows the accuracy and reliability of the model's prediction of the spectrum. The specific values ​​of the two parameter samples selected from the test set are shown in Table 2.

[0036] Table 2 Numerical values ​​of two parameter samples

[0037] 5. Use a spectrometer to test the DFB laser whose parameters are to be extracted, obtain the spectrum extraction information, and use it as the target spectrum extraction information; use the differential evolution algorithm to extract the spectrum parameters of the DFB laser, such as Figure 1 As shown, first, the population is initialized, and the initialized spectral parameters are input into the tested qualified model to predict the spectral extraction information, and the spectral extraction information output by the model is compared with the target spectral extraction information for error analysis. Then, the population is driven to evolve continuously through the process of mutation, crossover and selection, and the evolved parameters are input into the model to predict the spectral extraction information, and the spectral extraction information output by the model is compared with the target spectral extraction information for error analysis until the target accuracy or number of iterations is reached, and the extraction of spectral parameters is completed.

[0038] The target spectrum extraction information is the lasing spectrum of the DFB laser near the threshold tested by the spectrometer.

[0039] The differential evolution algorithm includes four steps: initialization, mutation, crossover, and selection. The specific process is as follows:

[0040] (1) Initialization For a 5-dimensional search space, population initialization is expressed as: ; in, Is the first in the G generation candidate solutions, each of which contains five spectral parameters, namely a set of values ​​of grating period, effective refractive index, phase, coupling coefficient and group refractive index. N is the size of the population. represents the value of the candidate solution in the jth dimension; Given the upper and lower bounds of the parameters, randomly initialize candidate solutions in the search space at Initial value of dimension ,in and In the The minimum and maximum values ​​of the search range for the dimension, is a uniform random number in the range of 0 to 1; After the population is initialized, the initialized spectral parameters are input into the deep learning model, and then the mean square error is used to calculate the error between the spectral extraction information of the actual test and the spectral extraction information predicted by the model;

[0041] (2) Variation The mutation operation is used to generate new candidate solutions: ; in, It is The mutation vector corresponding to the candidate solution is , and are three different candidate solutions in the population, and F is a scaling factor used to control the scale of the difference vector;

[0042] (3) Cross The mutated candidate solution and the original candidate solution are subjected to a binomial crossover operation to generate the test individuals, which can be expressed as: ; in, is the crossover probability, is a randomly selected dimension, ensuring that at least one dimension changes; When the evolutionary generation is G, The candidate solution The value of the dimension;

[0043] (4) Selection Finally, the selection operation between the mutated candidate solution and the original candidate solution is performed. The specific formula is as follows: ; in, is the objective function, Indicates the number of evolutionary generations in the parent population when G+1 Candidate solutions, Indicates the number of evolution generations in the crossover population when G A candidate solution, if If the fitness of is better, it will enter the next generation, otherwise the original candidate solution will be retained.

[0044] Table 3 lists the spectrum parameters of the laser extracted using the differential evolution algorithm. Based on these parameters, the wavelength coordinates and amplitudes of the spectrum peaks and valleys are predicted using the constructed model and compared with the information extracted from the target spectrum of the test. The results are shown in Figure 4 shown. Figure 5 The figure shows the comparison between the spectrum calculated by the transfer matrix method based on the extracted spectral parameters and the actual test spectrum. Figure 4 and Figure 5 It can be seen that the trends between the dotted line and the solid line are very consistent, indicating that the method combining deep learning and differential evolution algorithm can accurately and efficiently extract spectral parameters.

[0045] Table 3 Extraction parameter values

[0046] The present invention provides a method for extracting spectral parameters of a DFB laser based on the fusion of deep learning and differential evolution algorithms. A deep learning model is constructed by spectral parameters and spectral extraction information to characterize the mapping relationship of the transmission matrix, and the validity of the constructed model is evaluated by a test set. The population is initialized by the differential evolution algorithm, and the parameters are input into the constructed model to predict the spectral extraction information and the target spectral extraction information of the test for error analysis. Then, the population is driven to continuously evolve by the process of crossover, mutation and selection, and the spectral parameters are input into the constructed model to predict the spectral extraction information and the target spectral extraction information of the test for error analysis. When the error reaches the target value or the target number of iterations, the parameters are output to complete the parameter extraction. The invention extracts spectral parameters from the tested spectrum, and calculates the spectrum using the transmission matrix method based on the extracted spectral parameters, which is highly consistent with the tested spectrum. The invention can effectively avoid the need for a large number of numerical simulation operations when the traditional optimization algorithm extracts parameters, reduces the time cost, and can also realize the parameter extraction of batch DFB laser spectra.

[0047] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for extracting spectrum parameters of a DFB laser, characterized in that: The following steps are involved: Step 1, constructing a data set using spectral parameters and corresponding spectral extraction information, and constructing a deep learning model, using the spectral parameters as the input of the model, and the spectral extraction information as the output of the model, and using the data set to train and test the model; the spectral parameters include grating period, effective refractive index, coupling coefficient, phase and group refractive index; the spectral extraction information includes the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum; Step 2, using a spectrometer to test the DFB laser whose parameters are to be extracted, obtaining spectral extraction information, and using it as target spectral extraction information; using a differential evolution algorithm to extract the spectral parameters of the DFB laser, firstly initialize the population, and input the initialized spectral parameters into a qualified model to predict the spectral extraction information, and perform error analysis between the spectral extraction information output by the model and the target spectral extraction information, and then drive the population to continuously evolve through the process of mutation, crossover and selection, and input the evolved parameters into the model to predict the spectral extraction information, and then perform error analysis between the spectral extraction information output by the model and the target spectral extraction information, until the target accuracy or number of iterations is reached, and the extraction of spectral parameters is completed.

2. The method for extracting spectrum parameters of a DFB laser according to claim 1, characterized in that: In step 1, when constructing the data set, the range of the spectral parameters is first set, and then each parameter is uniformly selected based on the set range, and then the five spectral parameters are randomly combined to obtain several groups of spectral parameters. For each group of spectral parameters, the spectrum of the DFB laser is calculated using the transfer matrix method; According to the calculated spectrum, the wavelength coordinates and amplitude of the peak and valley values ​​of the spectrum are extracted to obtain the corresponding spectrum extraction information.

3. The method for extracting spectrum parameters of a DFB laser according to claim 1, characterized in that: In step 1, when constructing a data set, the spectral parameters are preprocessed by standardization, the spectral extraction information is preprocessed by maximum and minimum normalization, the preprocessed spectral parameters and spectral extraction information are used to construct a data set, and the data set is divided into a training set, a validation set, and a test set.

4. The method for extracting spectrum parameters of a DFB laser according to claim 1, characterized in that: In step 1, during the model training process, a loss function is used to quantify the difference between the spectral extraction information predicted by the model and the spectral extraction information in the training set; during the training process, the data in the validation set is used to measure whether the training process is overfitting; After the model training is completed, the model is tested using the data in the test set. If the test is qualified, proceed to the next step. If it fails, the model is retrained.

5. The method for extracting spectrum parameters of a DFB laser according to claim 1, characterized in that: In step 1, the constructed deep learning model includes an input layer, 5 hidden layers and an output layer; the input layer contains 5 neurons, corresponding to 5 spectral parameters respectively; the output layer contains 22 neurons, corresponding to the wavelength coordinates and amplitudes of 22 spectral peaks and valleys respectively; the number of neurons in the 5 hidden layers are 100, 200, 300, 200, and 100 respectively; batch normalization is introduced before each hidden layer to adjust the feature distribution and improve convergence stability.

6. A method for extracting spectrum parameters of a DFB laser according to claim 5, characterized in that: The activation function used by the deep learning model is the ReLu activation function.

7. The method for extracting spectrum parameters of a DFB laser according to claim 5, characterized in that: During the model training process, the loss functions used are mean square error and mean absolute error, and the formula is as follows: ； ； in, is the mean square error, is the mean absolute error, n is the total number of samples, It is The theoretical value of samples is It is The predicted value of a sample.

8. The method for extracting spectrum parameters of a DFB laser according to claim 5, characterized in that: The batch normalization method is as follows: In the training batches of dimensions There are A samples in is the first The first of the training dimensions samples; To standardize: ； in, For The data after standardization; is a very small constant used to prevent the denominator from being zero. For this training batch The mean of is the variance of the training batch; ； ； After scaling and translation factors: ； in, for The data after scaling and translation factor transformation is the input of the hidden layer. and denote the parameter vectors for scaling and translation respectively.

9. The method for extracting spectrum parameters of a DFB laser according to claim 1, characterized in that: In step 2, the target spectrum extraction information is the lasing spectrum of the DFB laser near the threshold tested by a spectrometer.

10. The method for extracting spectrum parameters of a DFB laser according to claim 1, characterized in that: In step 2, the differential evolution algorithm includes four steps: initialization, mutation, crossover, and selection. The specific process is as follows: (1) Initialization For a 5-dimensional search space, population initialization is expressed as: ； in, Is the first in the G generation candidate solutions, each of which contains five spectral parameters, namely a set of values ​​of grating period, effective refractive index, phase, coupling coefficient and group refractive index. N is the size of the population. represents the value of the candidate solution in the jth dimension; Initialize candidate solutions randomly in the search space Initial value of dimension ,in and In the The minimum and maximum values ​​of the search range for the dimension, is a uniform random number in the range of 0 to 1; After the population is initialized, the initialized spectral parameters are input into the deep learning model, and then the mean square error is used to calculate the error between the spectral extraction information of the actual test and the spectral extraction information predicted by the model; (2) Variation The mutation operation is used to generate new candidate solutions: ； in, It is The mutation vector corresponding to the candidate solution is , and are three different candidate solutions in the population, and F is a scaling factor used to control the scale of the difference vector; (3) Cross The mutated candidate solution and the original candidate solution are subjected to a binomial crossover operation to generate the test individuals, which can be expressed as: ； in, is the crossover probability, is a randomly selected dimension, ensuring that at least one dimension changes; When the evolutionary generation is G, The candidate solution The value of the dimension; (4) Selection Finally, the selection operation between the mutated candidate solution and the original candidate solution is performed. The specific formula is as follows: ； in, is the objective function, Indicates the number of evolutionary generations in the parent population when G+1 Candidate solutions, Indicates the number of evolution generations in the crossover population when G A candidate solution, if If the fitness of is better, it will enter the next generation, otherwise the original candidate solution will be retained.

Images