Multi-parameter automatic debugging method for optical modules based on deep learning

Abstract

The present invention discloses a method for automatic multi-parameter debugging of optical modules based on deep learning, which belongs to the field of optical communication technology. The method comprises the following steps: setting an adjustment parameter space of an optical module; constructing an optical module parameter debugging data set; constructing a deep neural network proxy model; training the deep neural network proxy model, and optimizing the optimal observation value set selected after training by using a Bayesian optimization acquisition model and a quasi-Newton method to obtain predicted performance parameters; judging whether the predicted performance parameters meet the sample addition condition, and if so, adding adversarial training samples to the optical module parameter debugging data set and retraining, otherwise saving the current optimal observation value set; repeating the training and saving the optimal observation value set until the optimal observation value set is separated by a certain time. t The optimal observation value sets saved in each round meet the convergence constraint, and the parameter debugging results are written into the Flash and stored. The present invention solves the problems of low efficiency and insufficient accuracy in optical module debugging.

Description

Technical Field

[0001] The present invention belongs to the field of optical communication technology, and in particular relates to a multi-parameter automatic debugging method for an optical module based on deep learning. Background Art

[0002] Currently, there are two main methods for parameter debugging and performance testing of optical modules. One is the DC power-on method, which directly adds a preset bias voltage to the optical module and tests its operating current and optical power. If they meet the set range, the optical module is qualified. However, this method tests a single parameter and cannot fully reflect the working status of the optical module. The other method is the manual debugging and testing method, which fine-tunes the optical module according to the product specification or chip manual, and uses an oscilloscope, spectrometer, optical power meter or other photoelectric sensor to determine whether the debugging or testing is qualified. Although this method improves the accuracy of debugging or testing, it greatly wastes time and is very inefficient, which is not conducive to mass production. Summary of the Invention

[0003] In response to the above-mentioned deficiencies in the prior art, the present invention provides a method for automatic multi-parameter debugging of optical modules based on deep learning, which solves the problems of low efficiency and insufficient accuracy during optical module debugging.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] The present invention provides a method for automatic multi-parameter debugging of an optical module based on deep learning, comprising the following steps:

[0006] S1. Set the adjustment parameter space of the optical module according to the target adjustment optical module test parameters;

[0007] S2. Obtain the optical module test parameters according to the adjustment parameter space of the optical module, and build an optical module parameter debugging data set;

[0008] S3, build a deep neural network agent model;

[0009] S4. Use the optical module parameter debugging data set to train the deep neural network proxy model, and optimize the optimal observation value set selected after training using the Bayesian optimization acquisition model and the quasi-Newton method to obtain the predicted performance parameters;

[0010] S5. Determine whether the predicted performance parameters meet the sample addition conditions. If so, add the adversarial training samples to the optical module parameter debugging dataset and return to S4. Otherwise, save the current optimal observation value set and enter S6.

[0011] S6. Repeat S4 to S5 for several rounds until the optimal observation value sets saved at intervals of τ rounds meet the convergence constraint. Then, write the current predicted performance parameters and their corresponding optical module test parameter samples into the Flash as parameter debugging results and store them.

[0012] The beneficial effects of the present invention are as follows: the present invention provides a method for automatic multi-parameter debugging of optical modules based on deep learning, which sets the adjustment parameter space of the optical module according to the target adjusted optical module test parameters, clarifies the optical module test parameters to be debugged, and provides a basis for obtaining the corresponding optical module test parameters and the optical module performance index parameters of interest to construct an optical module parameter debugging data set; the present invention constructs a deep neural network proxy model, establishes a mapping relationship between the optical module test parameters and the optical module performance index parameters, and realizes the quantification of prediction uncertainty by using the Monte Carlo random loss model for inference and random forward propagation sampling during training, which can effectively optimize and improve the depth Neural network proxy model; the present invention combines Bayesian optimization and deep neural network proxy models. Through the Bayesian optimization active learning strategy, the number of experiments is reduced, the testing cost is reduced, and the parameter space is efficiently explored to find the global optimal solution, that is, the performance parameters are predicted. Moreover, since the deep neural network proxy model can model complex nonlinear relationships and predict the performance of different parameter combinations, manual intervention is effectively reduced and the consistency and efficiency of the test are improved. The present invention realizes feedback control of the accuracy of the predicted performance parameters corresponding to the optical module test parameters by setting sample addition conditions, and ensures the optimality of the debugging parameters finally written to the Flash and stored by setting convergence constraints.

[0013] Other advantages of the present invention will be analyzed in more detail in subsequent embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0015] Figure 1 This is a flowchart of the steps of a method for automatic multi-parameter debugging of an optical module based on deep learning in an embodiment of the present invention.

[0016] Figure 2 Schematic diagram of the structure of the deep neural network proxy model in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, in one embodiment of the present invention, the present invention provides a method for automatic multi-parameter debugging of an optical module based on deep learning, comprising the following steps:

[0019] S1. Set the adjustment parameter space of the optical module according to the target adjustment optical module test parameters;

[0020] In this solution, the optical module test parameters included in the adjustment parameter space are bias current I B , modulation voltage V mod , TEC locking temperature T TEC , matching impedance Z0 and jitter rate jitter;

[0021] S2. Obtain the optical module test parameters according to the adjustment parameter space of the optical module, and build an optical module parameter debugging data set;

[0022] The S2 comprises the following steps:

[0023] S21. Using integrated automated test equipment to adjust the optical module test parameters in the adjustment parameter space in real time to test the optical module, and using a Latin hypercube sampling model to sample and time-series store the optical module test parameters and corresponding optical module performance indicator parameters to obtain an optical module parameter test data set;

[0024] The integrated automated test equipment in this solution is an integrated circuit automatic tester used in the semiconductor industry, used to test the integrity and quality of integrated circuit functions. By adjusting the optical module test parameters of the optical module under test, the integrated automated test equipment can obtain the corresponding performance indicators of the optical module under test. The optical module performance indicators in this solution include bit error rate (BER), extinction ratio (ER), optical power (P), and cross-point (CP).

[0025] The calculation expression of the Latin hypercube sampling model is as follows:

[0026] ,

[0027] in, Represents the optical module parameter test dataset, Indicates the test parameter samples of the i-th group of optical modules. represents the i-th group of optical module performance index parameter samples, and N represents the total number of parameter sample groups, where i = 1, 2, ..., N;

[0028] Latin hypercube sampling divides the range of each variable into several equally probable intervals, from which one sample value is selected for each variable. This ensures a uniform distribution of samples along each dimension, and Latin hypercube sampling can better cover the entire sample space. In this solution, the Latin hypercube sampling model can be used to obtain optical module test parameters that best represent optical module performance from limited samples, significantly reducing the problem of excessive sample data aggregation.

[0029] S22. Using optical simulation software to perform an optical module performance simulation test according to the optical module test parameters in the adjustment parameter space, and sampling and time-series storing the optical module test parameters and corresponding optical module performance index parameters in the simulation test to obtain an optical module parameter simulation data set;

[0030] S23. Combining the optical module parameter test data set and the optical module parameter simulation data set to obtain an initial optical module parameter debugging data set;

[0031] In this embodiment, the optical module test parameter sampling data set and the optical module simulation test parameter data set are stored through the HDF5 data model or the time series database to constitute the initial optical module parameter debugging data set; the HDF5 data model or the time series database can be used to trace the initial optical module parameter debugging data set by storing the version number, and can also detect abnormal data by setting deviation comparison, and add the abnormal data to the adversarial sample to update the optical module test parameter training set, providing a basis for improving the accuracy of deep neural networks.

[0032] S24. Performing Z-scale normalization processing on the optical module test parameters in the initial optical module parameter debugging data set to obtain normalized optical module test parameters;

[0033] The calculation expression of the normalized optical module test parameter is as follows:

[0034] , ,

[0035] in, It represents the normalized test parameter of the jth optical module. Indicates the jth optical module test parameter in the optical module parameter debugging data set. represents the mean value of the j-th optical module test parameter, represents the standard deviation of the jth optical module test parameter, and d represents the total number of optical module test parameters.

[0036] S25 . Replace the optical module test parameters in the optical module parameter debugging data set with the normalized optical module test parameters to obtain the optical module parameter debugging data set.

[0037] S3, build a deep neural network agent model;

[0038] like Figure 2 As shown, the deep neural network proxy model in S3 includes an input layer, a first fully connected layer, a random dropout layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, a fifth fully connected layer, and a mean output head and a standard deviation output head connected to the fifth fully connected layer, which are connected in sequence.

[0039] In this embodiment, the dimension of the training input data of the input layer is d; the first fully connected layer has a total of 512 nodes, and the training input data is feature integrated based on the Swish activation function and L2 regularization to obtain the first test data feature; the dropout rate of the random inactivation layer is 0.3, and after the first test data feature is input into the random inactivation layer, the first test data feature after random inactivation is obtained; the second fully connected layer has a total of 256 nodes, and the first test data feature is feature integrated based on the Swish activation function to obtain the second test data feature; the third fully connected layer has a total of 128 nodes, and the second test data feature is feature integrated based on the Swish activation function The third test data feature is integrated based on the row feature to obtain the third test data feature; the fourth fully connected layer has a total of 64 nodes, and the third test data feature is integrated based on the Swish activation function to obtain the fourth test data feature; the fifth fully connected layer has a total of 32 nodes, and the fourth test data feature is integrated based on the Swish activation function to obtain the fifth test data feature; the mean output head has a total of m nodes, and the fifth test data feature is linearly activated to obtain the predicted value of the optical module performance index parameter; the standard deviation output head has a total of m nodes, and the fifth test data feature is softplus activated to obtain the predicted noise of the optical module performance index parameter.

[0040] S4. Use the optical module parameter debugging data set to train the deep neural network proxy model, and optimize the optimal observation value set selected after training using the Bayesian optimization acquisition model and the quasi-Newton method to obtain the predicted performance parameters;

[0041] The S4 comprises the following steps:

[0042] S41, using the optical module test parameters in the optical module parameter debugging data set as training input data, and using the optical module performance index parameters corresponding to each optical module test data as the output true value, and dividing the optical module parameter debugging data set into a preset ratio as a training set;

[0043] S42. According to the multi-parameter joint loss function of the optical module, the deep neural network proxy model is repeatedly trained in preset batches using the training set to obtain a deep neural network proxy model that has completed stage training and a corresponding set of predicted values of optical module performance indicator parameters;

[0044] The S42 includes the following steps:

[0045] S421, inputting the training input data of dimension d in the training set into the input layer of the deep neural network proxy model, and using the first fully connected layer to perform feature integration on the training input data based on the Swish activation function and L2 regularization processing to obtain the first test data feature;

[0046] The calculation expression of the first test data feature is as follows:

[0047] ,

[0048] in, represents the first test data feature, represents the Swish activation function, represents the weight of the first fully connected layer, X represents the training input data, represents the bias of the first fully connected layer;

[0049] S422: Input the first test data feature into a random dropout layer to obtain the first test data feature after random dropout;

[0050] The calculation expression of the first test data feature after random deactivation is as follows:

[0051] ,

[0052] in, represents the first test data feature after random dropout, represents the random dropout layer, and p represents the dropout rate;

[0053] S423, using the second fully connected layer based on the Swish activation function to integrate the first test data features after random dropout to obtain second test data features;

[0054] The calculation expression of the second test data feature is as follows:

[0055] ,

[0056] in, represents the second test data feature, represents the weight of the second fully connected layer, represents the bias of the second fully connected layer;

[0057] S424. Using the third fully connected layer and the Swish activation function, perform feature integration on the second test data features to obtain third test data features.

[0058] The calculation expression of the third test data feature is as follows:

[0059] ,

[0060] in, represents the third test data feature, represents the weight of the third fully connected layer, represents the bias of the third fully connected layer;

[0061] S425. Using a fourth fully connected layer and a Swish activation function, perform feature integration on the third test data feature to obtain a fourth test data feature.

[0062] The calculation expression of the fourth test data feature is as follows:

[0063] ,

[0064] in, represents the fourth test data feature, represents the weight of the fourth fully connected layer, represents the bias of the fourth fully connected layer;

[0065] S426. Using a fifth fully connected layer, perform feature integration on the fourth test data feature based on a Swish activation function to obtain a fifth test data feature.

[0066] The calculation expression of the fifth test data feature is as follows:

[0067] ,

[0068] in, represents the fifth test data feature, represents the weight of the fifth fully connected layer, represents the bias of the fifth fully connected layer;

[0069] In this scheme, the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer and the fifth fully connected layer all use the Swish activation function, which can effectively alleviate the problem of gradient disappearance. The standard deviation output head uses the softplus activation function to ensure that the output is non-negative.

[0070] S427. Use the mean output head to perform linear activation on the fifth test data feature to obtain a predicted value of the optical module performance index parameter, and use the standard deviation output head to perform softplus activation on the fifth test data feature to obtain a predicted noise of the optical module performance index parameter.

[0071] The calculation expressions for the optical module performance index parameter prediction value and the optical module performance index parameter prediction noise are as follows:

[0072] ,

[0073] ,

[0074] in, represents the predicted value of the kth optical module performance index parameter, represents the mean output head weight, represents the mean output head bias, represents the predicted noise of the kth optical module performance parameter, Indicates the logarithmic value, e represents the exponential basis constant, represents the standard deviation output head weight, represents the standard deviation of the output head bias, where k = 1, 2, …, m, and m is the total number of optical module performance index parameters;

[0075] S428. According to the multi-parameter joint loss function of the optical module, the deep neural network proxy model is repeatedly trained in preset batches based on the method of S421-S427 to obtain a deep neural network proxy model that has completed stage training and a corresponding set of predicted values of optical module performance indicator parameters.

[0076] In this embodiment, when training the deep neural network proxy model, the weight attenuation coefficient in the multi-parameter joint loss function of the optical module is set to 10 -4 , which can effectively prevent data overfitting; in this embodiment, AdamW optimizer is selected and the initial learning rate lr is set to 10 -3 , and adopts a cosine annealing scheduler to flexibly select training data batches of 32~128 according to the real-time device memory size.

[0077] The calculation expression of the multi-parameter joint loss function of the optical module in S428 is as follows:

[0078] ,

[0079] ,

[0080] in, Represents the multi-parameter joint loss function of the optical module, represents the negative log-likelihood loss function, represents the L2 regularization coefficient, W represents the weight attenuation coefficient, It means to find the square of L2 norm, represents the predicted noise of the optical module performance index parameters corresponding to the i-th group of optical module test parameter samples, Indicates the predicted value of the optical module performance indicator parameter corresponding to the i-th group of optical module test parameter samples.

[0081] In this scheme, considering the joint influence of multiple uncertain parameters in the optical module, a multi-parameter joint loss function of the optical module is proposed based on the negative log-likelihood loss function, which effectively incorporates uncertainty into the training process. At the same time, the introduction of regularization terms ensures that the data will not be overfitted. The first term of the negative log-likelihood loss function in this scheme penalizes excessively high prediction uncertainty, effectively avoiding the deep neural network proxy model from being too conservative, while the second term of the negative log-likelihood loss function makes the predicted mean of the optical module performance index parameters close to the output true value, and when the error is large, it is compensated by high prediction noise.

[0082] The deep neural network proxy model uses a Monte Carlo random dropout model for inference and random forward propagation sampling during training;

[0083] The calculation expression of the Monte Carlo random dropout model is as follows:

[0084] ,

[0085] ,

[0086] in, It represents the predicted value of the optical module performance index parameter under Monte Carlo random dropout inference and forward propagation. It represents the predicted value of the optical module performance index parameter obtained by the t-th inference and forward propagation of the training input data. represents the prediction noise of the optical module performance index parameters under Monte Carlo random dropout inference and forward propagation, represents the predicted noise of the optical module performance parameter obtained by the t-th inference and forward propagation of the training input data, where t = 1, 2, …, T, and T represents the total number of random forward propagation samples.

[0087] In the present invention, the Monte Carlo random dropout model is used to quantify the uncertainty of the deep neural network proxy model's own predictions and the uncertainty caused by data noise, thereby providing a basis for optimizing and improving the deep neural network proxy model.

[0088] S43, performing a prediction deviation evaluation on the set of predicted values of the optical module performance indicator parameters to obtain a prediction deviation evaluation result;

[0089] The calculation expression of the prediction deviation evaluation result is as follows:

[0090] ,

[0091] ,

[0092] in, represents the prediction deviation evaluation result, Indicates the true value of the performance index parameter of the vth group of optical modules in the set of predicted values of the performance index parameter of the optical module. Indicates the predicted value of the performance index parameter of the vth group of optical modules in the set of predicted values of the performance index parameter of the optical module. represents the average predicted value of the optical module performance index parameter in the optical module performance index parameter predicted value set, where v=1, 2, …, V, where V represents the total number of groups of optical module performance index parameter predicted values in the optical module performance index parameter predicted value set;

[0093] In this solution, the prediction deviation evaluation result is used to calculate the deviation between the predicted value of the optical module performance index parameter in the optical module performance index parameter prediction value set and its corresponding true value. The larger the value of the prediction deviation evaluation result, the closer the predicted value is to the true value, and the better the performance of the deep neural network proxy model. The solution of the present invention ensures the effectiveness of the deep neural network proxy model training by setting up regular monitoring of the prediction deviation evaluation result, adjusting the model parameters and adding adversarial training samples for retraining when the prediction deviation evaluation result is negative.

[0094] S44. If the prediction deviation result is negative, adjust the learning rate and training data batches for training the deep neural network proxy model, add adversarial training samples, and continue to retrain the deep neural network proxy model and evaluate the prediction deviation based on the method of S42-S43;

[0095] S45. If the prediction deviation result is not negative, after repeatedly training the deep neural network proxy model and evaluating the prediction deviation for a preset number of iterations based on the method of S42-S43, the set of predicted values of the optical module performance index parameters corresponding to the maximum prediction deviation evaluation result is selected as the optimal observation value set;

[0096] S46. Based on the optimal observation value set, the expected improved optimization result is calculated by using the Bayesian optimization acquisition model;

[0097] The calculation expression of the Bayesian optimization acquisition model is as follows:

[0098] ,

[0099] in, Indicates the expected improvement optimization result corresponding to the optical module test parameter sample X. Expressing hope, Indicates the predicted value of the optical module performance index parameter corresponding to the optical module test parameter sample X. represents the optimal set of observations, It represents the predicted mean value of the optical module performance index parameter corresponding to the optical module test parameter sample X. Indicates taking the non-negative maximum value, represents the exploration weight factor;

[0100] S47, performing boundary constraint optimization on the expected improvement optimization result by using the quasi-Newton method L-BFGS-B to obtain the predicted performance parameters;

[0101] The calculation expression is as follows:

[0102] ,

[0103] in, represents the prediction performance parameter, Indicates taking the optical module test parameter sample corresponding to the maximum expected improvement optimization result. Indicates that, The constraint function of the optical module test parameter sample is represented. In this embodiment, the constraint function of the optical module test sample includes, for example, a bias current range greater than or equal to 10 mA and less than or equal to 80 mA, and an optical power less than 4 mW.

[0104] S5. Determine whether the predicted performance parameters meet the sample addition conditions. If so, add the adversarial training samples to the optical module parameter debugging dataset and return to S4. Otherwise, save the current optimal observation value set and enter S6.

[0105] The calculation expression of the sample addition condition in S5 is as follows:

[0106] ,

[0107] in, Indicates the optical module test parameter samples corresponding to the predicted performance parameters. Indicates the optical module performance index parameter sample corresponding to the predicted performance parameter. Indicates the prediction noise of the optical module performance indicator parameter sample corresponding to the predicted performance parameter;

[0108] The adversarial training samples in S5 are the predicted performance parameters and the optical module test parameter samples corresponding to the predicted performance parameters.

[0109] S6. Repeat S4 to S5 for several rounds until the optimal observation value sets saved at intervals of τ rounds meet the convergence constraint. Then, write the current predicted performance parameters and their corresponding optical module test parameter samples into the Flash as parameter debugging results and store them.

[0110] The calculation expression of the convergence constraint in S6 is as follows:

[0111] ,

[0112] in, represents the optimal set of observations saved in the current round, represents the optimal set of observations saved τ rounds ago, Indicates the absolute value. Indicates the convergence threshold. In this embodiment, the convergence threshold is set to .

[0113] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for automatic multi-parameter debugging of optical modules based on deep learning, characterized in that: The steps include: S1. Set the adjustment parameter space of the optical module according to the target adjustment optical module test parameters; S2. Obtain the optical module test parameters according to the adjustment parameter space of the optical module, and build an optical module parameter debugging data set; The S2 comprises the following steps: S21. Using integrated automated test equipment to adjust the optical module test parameters in the adjustment parameter space in real time to test the optical module, and using a Latin hypercube sampling model to sample and time-series store the optical module test parameters and corresponding optical module performance indicator parameters to obtain an optical module parameter test data set; The calculation expression of the Latin hypercube sampling model is as follows: ， in, Represents the optical module parameter test dataset, Indicates the test parameter samples of the i-th group of optical modules. represents the i-th group of optical module performance index parameter samples, and N represents the total number of parameter sample groups, where i = 1, 2, ..., N; S22. Using optical simulation software to perform an optical module performance simulation test according to the optical module test parameters in the adjustment parameter space, and sampling and time-series storing the optical module test parameters and corresponding optical module performance index parameters in the simulation test to obtain an optical module parameter simulation data set; S23. Combining the optical module parameter test data set and the optical module parameter simulation data set to obtain an initial optical module parameter debugging data set; S24. Performing Z-scale normalization processing on the optical module test parameters in the initial optical module parameter debugging data set to obtain normalized optical module test parameters; S25. Replace the optical module test parameters in the optical module parameter debugging data set with the normalized optical module test parameters to obtain the optical module parameter debugging data set; S3, build a deep neural network agent model; S4. Use the optical module parameter debugging data set to train the deep neural network proxy model, and optimize the optimal observation value set selected after training using the Bayesian optimization acquisition model and the quasi-Newton method to obtain the predicted performance parameters; The S4 comprises the following steps: S41, using the optical module test parameters in the optical module parameter debugging data set as training input data, and using the optical module performance index parameters corresponding to each optical module test data as the output true value, and dividing the optical module parameter debugging data set into a preset ratio as a training set; S42. According to the multi-parameter joint loss function of the optical module, the deep neural network proxy model is repeatedly trained in preset batches using the training set to obtain a deep neural network proxy model that has completed stage training and a corresponding set of predicted values of optical module performance indicator parameters; S43, performing a prediction deviation evaluation on the set of predicted values of the optical module performance indicator parameters to obtain a prediction deviation evaluation result; The calculation expression of the prediction deviation evaluation result is as follows: ， ， in, represents the prediction deviation evaluation result, Indicates the true value of the performance index parameter of the vth group of optical modules in the set of predicted values of the performance index parameter of the optical module. Indicates the predicted value of the performance index parameter of the vth group of optical modules in the set of predicted values of the performance index parameter of the optical module. represents the average predicted value of the optical module performance index parameter in the optical module performance index parameter predicted value set, where v=1, 2, …, V, where V represents the total number of groups of optical module performance index parameter predicted values in the optical module performance index parameter predicted value set; S44. If the prediction deviation result is negative, adjust the learning rate and training data batches for training the deep neural network proxy model, add adversarial training samples, and continue to retrain the deep neural network proxy model and evaluate the prediction deviation based on the method of S42-S43; S45. If the prediction deviation result is not negative, after repeatedly training the deep neural network proxy model and evaluating the prediction deviation for a preset number of iterations based on the method of S42-S43, the set of predicted values of the optical module performance index parameters corresponding to the maximum prediction deviation evaluation result is selected as the optimal observation value set; S46. Based on the optimal observation value set, the expected improved optimization result is calculated by using the Bayesian optimization acquisition model; The calculation expression of the Bayesian optimization acquisition model is as follows: ， in, Indicates the expected improvement optimization result corresponding to the optical module test parameter sample X. Expressing hope, Indicates the predicted value of the optical module performance index parameter corresponding to the optical module test parameter sample X. represents the optimal set of observations, It represents the predicted mean value of the optical module performance index parameter corresponding to the optical module test parameter sample X. Indicates taking the non-negative maximum value, represents the exploration weight factor; S47, performing boundary constraint optimization on the expected improvement optimization result by using the quasi-Newton method L-BFGS-B to obtain the predicted performance parameters; The calculation expression is as follows: ， in, represents the prediction performance parameter, Indicates taking the optical module test parameter sample corresponding to the maximum expected improvement optimization result. Indicates that, A constraint function representing a sample of optical module test parameters; S5. Determine whether the predicted performance parameters meet the sample addition conditions. If so, add the adversarial training samples to the optical module parameter debugging dataset and return to S4. Otherwise, save the current optimal observation value set and enter S6. S6. Repeat S4 to S5 for several rounds until the optimal observation value sets saved at intervals of τ rounds meet the convergence constraint. Then, write the current predicted performance parameters and their corresponding optical module test parameter samples into the Flash as parameter debugging results and store them.

2. The optical module multi-parameter automatic debugging method based on deep learning according to claim 1 is characterized in that: The deep neural network proxy model in S3 includes an input layer, a first fully connected layer, a random inactivation layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, a fifth fully connected layer, and a mean output head and a standard deviation output head connected to the fifth fully connected layer.

3. The optical module multi-parameter automatic debugging method based on deep learning according to claim 2 is characterized in that: The S42 includes the following steps: S421, inputting the training input data of dimension d in the training set into the input layer of the deep neural network proxy model, and using the first fully connected layer to perform feature integration on the training input data based on the Swish activation function and L2 regularization processing to obtain the first test data feature; S422: Input the first test data feature into a random dropout layer to obtain the first test data feature after random dropout; S423, using the second fully connected layer based on the Swish activation function to integrate the first test data features after random dropout to obtain second test data features; S424. Using the third fully connected layer and the Swish activation function, perform feature integration on the second test data features to obtain third test data features. S425. Using a fourth fully connected layer and a Swish activation function, perform feature integration on the third test data feature to obtain a fourth test data feature. S426. Using a fifth fully connected layer, perform feature integration on the fourth test data feature based on a Swish activation function to obtain a fifth test data feature. S427. Use the mean output head to perform linear activation on the fifth test data feature to obtain a predicted value of the optical module performance index parameter, and use the standard deviation output head to perform softplus activation on the fifth test data feature to obtain a predicted noise of the optical module performance index parameter. S428. According to the multi-parameter joint loss function of the optical module, the deep neural network proxy model is repeatedly trained in preset batches based on the method of S421-S427 to obtain a deep neural network proxy model that has completed stage training and a corresponding set of predicted values of optical module performance indicator parameters.

4. The optical module multi-parameter automatic debugging method based on deep learning according to claim 3 is characterized in that: The calculation expression of the multi-parameter joint loss function of the optical module in S428 is as follows: ， ， in, represents the multi-parameter joint loss function of the optical module, represents the negative log-likelihood loss function, represents the L2 regularization coefficient, W represents the weight attenuation coefficient, It means to find the square of L2 norm, represents the predicted noise of the optical module performance index parameters corresponding to the i-th group of optical module test parameter samples, Indicates the predicted value of the optical module performance indicator parameter corresponding to the i-th group of optical module test parameter samples.

5. The optical module multi-parameter automatic debugging method based on deep learning according to claim 4 is characterized in that: The deep neural network proxy model uses a Monte Carlo random dropout model for inference and random forward propagation sampling during training; The calculation expression of the Monte Carlo random dropout model is as follows: ， ， in, It represents the predicted value of the optical module performance index parameter under Monte Carlo random dropout inference and forward propagation. It represents the predicted value of the optical module performance index parameter obtained by the t-th inference and forward propagation of the training input data. represents the prediction noise of the optical module performance index parameters under Monte Carlo random dropout inference and forward propagation, represents the predicted noise of the optical module performance parameter obtained by the t-th inference and forward propagation of the training input data, where t = 1, 2, …, T, and T represents the total number of random forward propagation samples.

6. The optical module multi-parameter automatic debugging method based on deep learning according to claim 1 is characterized in that: The calculation expression of the sample addition condition in S5 is as follows: ， in, Indicates the optical module test parameter samples corresponding to the predicted performance parameters. Indicates the optical module performance index parameter sample corresponding to the predicted performance parameter. Indicates the prediction noise of the optical module performance indicator parameter sample corresponding to the predicted performance parameter; The adversarial training samples in S5 are the predicted performance parameters and the optical module test parameter samples corresponding to the predicted performance parameters.

7. The optical module multi-parameter automatic debugging method based on deep learning according to claim 1 is characterized in that: The calculation expression of the convergence constraint in S6 is as follows: ， in, represents the optimal set of observations saved in the current round, represents the optimal set of observations saved τ rounds ago, Indicates the absolute value. Indicates the convergence threshold.

Images