Optical communication system sample dynamic modeling method and system
By building a simulation system in an optical communication system, collecting data and dynamically selecting sampling points to train a surrogate model, the problems of low model training efficiency and high cost in existing technologies are solved, achieving high-precision and low-cost modeling results.
Patent Information
- Application Number
- CN202511742884.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-03
AI Technical Summary
In the training process of existing optical communication system models, the limited datasets are not effectively utilized, resulting in low efficiency in improving model performance in complex and ever-changing optical communication scenarios, high modeling costs, and an inability to specifically select data points in key areas for training.
By building an optical communication simulation system, collecting simulation and experimental data, initializing the surrogate model, establishing a multi-fidelity prediction acquisition function, dynamically selecting sampling points, progressively augmenting the dataset, training the surrogate model, achieving adaptive learning, and finally obtaining a model with satisfactory accuracy.
It reduces the scale of experimental data required for training, lowers modeling costs, improves the accuracy and efficiency of the model in complex scenarios, and significantly reduces the dependence on expensive experimental data.
Smart Images
Figure CN121603815A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of simulation design technology, specifically to a method and system for dynamic modeling of optical communication system samples, and more particularly to a method for dynamic modeling of optical communication systems with small samples. Background Technology
[0002] An existing paper (JW Nevin, FJ Vaquero-Caballero, DJ Ives and S. J. Savory, "Physics-Informed Gaussian Process Regression for Optical Fiber Communication Systems," in Journal of Lightwave Technology, vol. 39, no. 21, pp. 6833-6844, 1 Nov. 1, 2021) mentions incorporating prior information from the physical model into the training of the Gaussian process to reduce the amount of labeled data required. However, during method validation, it did not consider evaluating the potential sampling space in complex and variable optical communication scenarios, nor did it dynamically select the dataset more specifically by progressively increasing the training sample points. This leaves the model unaware of which regions in the input space are high-uncertainty areas and which regions the physical model might fail; and the lack of dynamic dataset selection through progressively increasing training sample points prevents the training process from efficiently improving the model's performance in critical regions. The training data remains random or pre-set, and data points in the boundary or highly nonlinear regions that have the greatest impact on system performance may still be sparse, failing to make targeted use of limited and expensive measurement data. This necessitates more experimental data during model training, increasing modeling costs, otherwise performance will degrade. Summary of the Invention
[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for dynamic modeling of optical communication system samples.
[0004] A dynamic modeling method for optical communication system samples provided by the present invention includes the following steps: Step S1: Build an optical communication simulation system and corresponding experimental environment; collect simulation data and experimental data from the optical communication simulation system and experimental environment; initialize the proxy model using the simulation data and experimental data; establish performance evaluation indicators to evaluate the validity of the data and the model's compliance. Step S2: Establish a multi-fidelity predictive acquisition function, use a surrogate model to evaluate potential sampling points in the input space, and determine the subsequent sampling strategy based on the evaluation results through the multi-fidelity predictive acquisition function; Step S3: Re-collect data using the determined sampling strategy, augment the re-collected data points into the dataset, and retrain the surrogate model using the dataset to achieve continuous adaptive learning that matches the global accuracy of the model, thus obtaining an optimized surrogate model. Step S4: Repeat steps S2 to S3, evaluate the model's compliance based on the performance evaluation indicators established in step S1, and finally obtain a proxy model with satisfactory accuracy.
[0005] Preferably, the sampling method for the initialization data is random sampling or Latin hypercube sampling.
[0006] Preferably, the proxy model is a serial co-kickinic proxy model, expressed as:
[0007] in, It is any combination of parameters in the input space, corresponding to the adjustable variables of the optical communication simulation system; It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link; It is a low-fidelity function that provides an estimate of a given performance metric, based on a physical model; scaling factor measure right The global linear contribution is determined through the optimization process; the residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent of each other.
[0008] Preferably, in step S1, the modeling process of the proxy model is as follows: Collect sample points to obtain simulation data and experimental data ; The two components of the proxy model are modeled using independent Gaussian processes, and modeled using simulation data. Modeling using the difference between experimental and simulation data ; Using the complete dataset All hyperparameters in the surrogate model are jointly optimized using maximum likelihood estimation; hyperparameters include: kernel function parameters of the Gaussian process, scaling factor, etc. ; The trained proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance .
[0009] Preferably, in step S2, the multi-fidelity prediction acquisition function is used to balance information gain and sampling cost, and its expression is as follows:
[0010] in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance; The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.
[0011] The present invention also provides a dynamic modeling system for optical communication system samples, comprising the following modules: Module M1: Build an optical communication simulation system and corresponding experimental environment; collect simulation data and experimental data from the optical communication simulation system and experimental environment; initialize the proxy model using the simulation data and experimental data; establish performance evaluation indicators to evaluate the validity of the data and the model's compliance. Module M2: Establishes a multi-fidelity predictive acquisition function, uses a surrogate model to evaluate potential sampling points in the input space, and determines subsequent sampling strategies based on the evaluation results through the multi-fidelity predictive acquisition function; Module M3: Data is re-collected using a defined sampling strategy, the re-collected data points are augmented into the dataset, and the surrogate model is retrained using the dataset to achieve continuous adaptive learning that matches the global accuracy of the model, resulting in an optimized surrogate model. Module M4: Repeatedly executes modules M2 to M3, evaluates the model's compliance based on the performance evaluation indicators established in module M1, and finally obtains a proxy model that meets the accuracy requirements.
[0012] Preferably, the sampling method for the initialization data is random sampling or Latin hypercube sampling.
[0013] Preferably, the proxy model is a serial co-kickinic proxy model, expressed as:
[0014] in, It is any combination of parameters in the input space, corresponding to the adjustable variables of the optical communication simulation system; It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link; It is a low-fidelity function that provides an estimate of a given performance metric, based on a physical model; scaling factor measure right The global linear contribution is determined through the optimization process; the residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent of each other.
[0015] Preferably, in module M1, the modeling process of the proxy model is as follows: Collect sample points to obtain simulation data and experimental data ; The two components of the proxy model are modeled using independent Gaussian processes, and modeled using simulation data. Modeling using the difference between experimental and simulation data ; Using the complete dataset All hyperparameters in the surrogate model are jointly optimized using maximum likelihood estimation; hyperparameters include: kernel function parameters of the Gaussian process, scaling factor, etc. ; The trained proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance .
[0016] Preferably, in module M2, the multi-fidelity prediction acquisition function is used to balance information gain and sampling cost, and its expression is as follows:
[0017] in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance; The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.
[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. In establishing the multi-fidelity prediction acquisition function, this invention separates the low-fidelity model based on the physical model from its nonlinear residuals with the real experimental data, and establishes corresponding Gaussian processes (providing the mean and variance for each set of inputs) to evaluate the uncertainty of each potential sampling point. The selection of sample points is carried out generationally, and each time the current training model evaluates potential sampling points in the sampling space to find the point that makes the multi-fidelity prediction acquisition function reach its maximum value, and incorporates it into the next data acquisition. Since the influence of different regions of the sampling space on model performance in specific scenarios is considered, the selection of sampling points can be more targeted, rather than random or uniform sampling across dimensions. This reduces the scale of experimental data required for training, effectively reduces modeling costs, and is therefore applicable to more complex and variable optical communication scenarios, enabling the establishment of more accurate models at a lower cost.
[0019] 2. This invention establishes a multi-fidelity Gaussian process model and evaluates the uncertainty of each potential sampling point, enabling the model to clearly identify high uncertainty regions in the input space and boundary regions where the physical model may fail, thus achieving transparent quantification of the model's knowledge boundaries.
[0020] 3. By integrating inexpensive low-fidelity physical model information with high-value high-fidelity experimental data and combining it with a dynamic point selection strategy, this invention significantly reduces the scale of expensive experimental data required to achieve the target accuracy, effectively reduces the modeling cost of optical communication systems, and ensures the accuracy of the model in highly complex scenarios.
[0021] 4. This invention solves the problem of inaccurate low-fidelity data and high cost of high-fidelity data acquisition by adopting a serial co-kriging modeling method, and achieves the effect of effectively guiding experimental data acquisition with simulation data.
[0022] 5. This invention solves the problem of insufficient accuracy when modeling with small sample datasets by adopting a method that combines residual uncertainty with dynamic point selection, and achieves the effect of bringing the maximum information gain in each experiment. Attached Figure Description
[0023] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of the optical communication system sample dynamic modeling method of the present invention. Detailed Implementation
[0024] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0025] Example 1 like Figure 1 As shown, this embodiment provides a method for dynamic modeling of optical communication system samples, including the following steps: Step S1: Build an optical communication simulation system and corresponding experimental environment; collect simulation data and experimental data from the optical communication simulation system and experimental environment; initialize the proxy model using the simulation data and experimental data; establish performance evaluation indicators to evaluate the validity of the data and the model's compliance. The initialization data is sampled using random sampling or Latin hypercube sampling. The proxy model is a serial co-kickinic proxy model, expressed as:
[0026] in, It is any combination of parameters in the input space, corresponding to the adjustable variables of the optical communication simulation system; It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link; It is a low-fidelity function that provides an estimate of a given performance metric, based on a physical model; scaling factor measure right The global linear contribution is determined through the optimization process; the residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent; The specific modeling process for the proxy model is as follows: Collect sample points to obtain simulation data and experimental data ; The two components of the proxy model are modeled using independent Gaussian processes, and modeled using simulation data. Modeling using the difference between experimental and simulation data ; Using the complete dataset All hyperparameters in the surrogate model are jointly optimized using maximum likelihood estimation; hyperparameters include: kernel function parameters of the Gaussian process, scaling factor, etc. ; The trained proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance .
[0027] mean It is a proxy model for the input correspondence High-fidelity objective function The optimal mean estimate of the true value, provided by the surrogate model based on existing data and a probabilistic framework. Most likely values; variance It is a proxy model prediction The uncertainty variance over time reflects the reliability of the prediction results and measures the surrogate model's reliability. The range of guesses for the true value: the larger the variance, the better the surrogate model's accuracy. The more uncertain the true value, the smaller the variance, and the more concentrated the prediction results, the better the surrogate model is. The higher the credibility of the corresponding true value.
[0028] Step S2: Establish a multi-fidelity predictive acquisition function, use a surrogate model to evaluate potential sampling points in the input space, and determine the subsequent sampling strategy based on the evaluation results through the multi-fidelity predictive acquisition function; The multi-fidelity prediction acquisition function is used to balance information gain and sampling cost, and its expression is as follows:
[0029] in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance; The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.
[0030] It can be translated as expected variance reduction, which measures "if at the input point..." Supplement the target function for data collection. The quantitative indicator is "how much the overall prediction variance of the model can be reduced based on the corresponding sample data".
[0031] Step S3: Re-collect data using the determined sampling strategy, augment the re-collected data points into the dataset, and retrain the surrogate model using the dataset to achieve continuous adaptive learning that matches the global accuracy of the model, thus obtaining an optimized surrogate model.
[0032] Step S4: Repeat steps S2 to S3, evaluate the model's compliance based on the performance evaluation indicators established in step S1, and finally obtain a proxy model with satisfactory accuracy.
[0033] The proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance Therefore, in each iteration, the trained multi-fidelity model can be used to evaluate all potential, uncollected sampling points x in the entire input sampling space, and then utilize... The point at which the multi-fidelity prediction acquisition function reaches its maximum value is calculated and incorporated into the next step of data acquisition.
[0034] Furthermore, in step S4, the process of repeating steps S2 to S3 is as follows: The optimized surrogate model from step S3 is then applied back to step S2 to re-evaluate the potential sampling points in the input space. The optimal sampling points obtained in step S2 are then experimentally sampled and reintroduced into the dataset. The model is then trained on the updated dataset.
[0035] This loop mechanism efficiently maximizes information gain, aiming to accurately identify the regions in the entire input space where the current surrogate model is most uncertain or has the greatest potential for information gain. By maximizing the sampling function, the model avoids expensive sampling in regions where sufficient information is already available, thus avoiding resource waste caused by random or uniform sampling.
[0036] Throughout the iterative convergence process, initially, the model data is sparse, and the variance is high in most areas; step S2 tends to explore large blank areas. In the middle stage, the model accuracy gradually improves, and high-variance regions begin to concentrate at the boundaries of the input space or in highly nonlinear and complex regions; step S2 at this point becomes more targeted, focusing on these challenging regions. Later, as the number of iterations increases, the model's prediction variance across the entire input space continuously decreases, indicating that the model's uncertainty is gradually eliminated. Finally, when the uncertainty at all potential sampling points is sufficiently small or the evaluation metrics meet the requirements, it means that the model has a sufficiently accurate and high-confidence understanding of the entire sampling space, i.e., the accuracy target has been achieved.
[0037] Furthermore, the evaluation metric in step S4 is: mean squared error (the points in the sampling space are generally lower than a given mean squared error value through the model's output, such as 0.5 dB). Alternatively, the maximum number of sampling iterations or the upper limit of the sampling dataset size can be used to terminate the iteration process.
[0038] The present invention also provides a dynamic modeling system for optical communication system samples. The dynamic modeling system for optical communication system samples can be implemented by executing the process steps of the dynamic modeling method for optical communication system samples. That is, those skilled in the art can understand the dynamic modeling method for optical communication system samples as a preferred embodiment of the dynamic modeling system for optical communication system samples.
[0039] Example 2 This embodiment provides a dynamic modeling system for optical communication system samples, including the following modules: Module M1: Build an optical communication simulation system and corresponding experimental environment; collect simulation data and experimental data from the optical communication simulation system and experimental environment; initialize the proxy model using the simulation data and experimental data; establish performance evaluation indicators to evaluate the validity of the data and the model's compliance. The initialization data is sampled using random sampling or Latin hypercube sampling. The proxy model is a serial co-kickinic proxy model, expressed as:
[0040] in, It is any combination of parameters in the input space, corresponding to the adjustable variables of the optical communication simulation system; It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link; It is a low-fidelity function that provides an estimate of a given performance metric, based on a physical model; scaling factor measure right The global linear contribution is determined through the optimization process; the residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent; The specific modeling process for the proxy model is as follows: Collect sample points to obtain simulation data and experimental data ; The two components of the proxy model are modeled using independent Gaussian processes, and modeled using simulation data. Modeling using the difference between experimental and simulation data ; Using the complete dataset All hyperparameters in the surrogate model are jointly optimized using maximum likelihood estimation; hyperparameters include: kernel function parameters of the Gaussian process, scaling factor, etc. ; The trained proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance .
[0041] Module M2: Establishes a multi-fidelity predictive acquisition function, uses a surrogate model to evaluate potential sampling points in the input space, and determines subsequent sampling strategies based on the evaluation results through the multi-fidelity predictive acquisition function; The multi-fidelity prediction acquisition function is used to balance information gain and sampling cost, and its expression is as follows:
[0042] in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance; The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.
[0043] Module M3: Data is re-collected using a defined sampling strategy, the re-collected data points are augmented into the dataset, and the surrogate model is retrained using the dataset to achieve continuous adaptive learning that matches the global accuracy of the model, resulting in an optimized surrogate model. Module M4: Repeatedly executes modules M2 to M3, evaluates the model's compliance based on the performance evaluation indicators established in module M1, and finally obtains a proxy model that meets the accuracy requirements.
[0044] The proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance Therefore, in each iteration, the trained multi-fidelity model can be used to evaluate all potential, uncollected sampling points x in the entire input sampling space, and then utilize... The point at which the multi-fidelity prediction acquisition function reaches its maximum value is calculated and incorporated into the next step of data acquisition.
[0045] Furthermore, in module M4, the process of repeatedly executing modules M2 to M3 is as follows: The optimized surrogate model from module M3 is then applied to module M2 to re-evaluate the potential sampling points in the input space. The optimal sampling points obtained from the evaluation in module M2 are then experimentally sampled and reintroduced into the dataset. The model is then trained on the updated dataset.
[0046] This loop mechanism efficiently maximizes information gain, aiming to accurately identify the regions in the entire input space where the current surrogate model is most uncertain or has the greatest potential for information gain. By maximizing the sampling function, the model avoids expensive sampling in regions where sufficient information is already available, thus avoiding resource waste caused by random or uniform sampling.
[0047] Throughout the iterative convergence process, initially, the model data is sparse, and variance is high in most areas; module M2 tends to explore large blank areas. In the middle stage, model accuracy gradually improves, and high-variance regions begin to concentrate at the boundaries of the input space or in highly nonlinear and complex regions; module M2's sampling becomes more targeted at this point, focusing on these challenging regions. In the later stage, as the number of iterations increases, the model's prediction variance across the entire input space continuously decreases, indicating that model uncertainty is gradually eliminated. Finally, when the uncertainty at all potential sampling points is sufficiently small or the evaluation metrics meet the requirements, it means that the model has a sufficiently accurate and high-confidence understanding of the entire sampling space, i.e., accuracy has been achieved.
[0048] Furthermore, the evaluation metrics in module M4 are: mean squared error (the points in the sampling space are generally lower than a given mean squared error value through the model's output, such as 0.5 dB), and the iteration process can also be terminated by using the maximum number of sampling iterations or the upper limit of the sampling dataset size.
[0049] Example 3 Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.
[0050] This embodiment employs a surrogate model framework based on a dynamic sampling strategy to achieve high-precision, low-cost modeling of optical communication systems. This embodiment provides a small-sample dynamic modeling method for optical communication systems, the specific technical solution of which is as follows: Step 1: Build an optical communication simulation system and corresponding experimental environment, establish performance evaluation indicators, collect a small amount of simulation and experimental data, and initialize the proxy model; Step 2: Establish a multi-fidelity predictive acquisition function, use the existing surrogate model to evaluate potential sampling points in the input space, and determine the next sampling strategy; Step 3: Augment the newly collected data points into the dataset, retrain the existing proxy model, and achieve continuous adaptive learning that matches the global accuracy of the model; Step 4: Repeat steps 2-3 to finally obtain a proxy model with satisfactory accuracy.
[0051] Furthermore, "small sample" means that, in order to establish a model with the same data accuracy, the amount of data required by this invention is much smaller than the amount of data required for modeling through random sampling or uniform sampling. Generally speaking, uniform sampling requires hundreds to thousands of samples, but this invention requires only tens or just over a hundred samples.
[0052] Furthermore, in step 1, the sampling method for the initialization data includes, but is not limited to, random sampling and Latin hypercube sampling.
[0053] Furthermore, in step 1, the modeling method for the proxy model is a serial co-kriging proxy model, the expression of which is:
[0054] In the formula It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link. It is a low-fidelity function that provides a fast but biased estimate of a given performance metric, based on a physical model. Scaling factor measure right The global linear contribution is determined through the optimization process. Residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent of each other.
[0055] Furthermore, the modeling process in step 1 is as follows: A small number of sample points are collected using methods such as random sampling to obtain simulation data. and experimental data ; The proxy model consists of two independent Gaussian processes. Modeling is done using simulation data. Modeling using the difference between experimental and simulation data ; Utilizing the current complete dataset All hyperparameters in the joint optimization model are estimated using maximum likelihood estimation. Hyperparameters include: the kernel function parameter of the Gaussian process, scaling factor, etc. ; Once training is complete, the existing model can provide any input point. place Complete posterior probability distribution, i.e., predicted mean and variance .
[0056] Furthermore, in step 2, the multi-fidelity prediction acquisition function is used to balance information gain and sampling cost. Its expression is as follows:
[0057] in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance. The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.
[0058] To reduce the cost of collecting low-fidelity data, To reduce the cost of collecting high-fidelity data, The model will first be Larger areas are densely sampled using simulation. However, once the benefits of low-fidelity data reach saturation, further effective reduction becomes impossible. Instead, they opted for more expensive high-fidelity data sampling.
[0059] It is only considered that the cost of high-fidelity data is at least K times that of low-fidelity data. It is valid. K is usually between 10 and 100, and can be specifically determined by the ratio of the physical model running time required for the specific scenario to the data acquisition time required for the experiment.
[0060] Furthermore, in step 4, once the current proxy model reaches the target accuracy on the test set, sampling training is stopped.
[0061] This invention proposes a small-sample dynamic modeling method for optical communication systems, which deeply integrates sequence sampling with surrogate models, quantifies the uncertainty of the potential sampling space, iteratively selects points and trains generation by generation, and achieves cost optimization across fidelity scenarios.
[0062] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.
[0063] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A method for dynamic modeling of samples in an optical communication system, characterized in that, Includes the following steps: Step S1: Build an optical communication simulation system and corresponding experimental environment; collect simulation data and experimental data from the optical communication simulation system and experimental environment; initialize the proxy model using the simulation data and experimental data; establish performance evaluation indicators to evaluate the validity of the data and the model's compliance. Step S2: Establish a multi-fidelity predictive acquisition function, use a surrogate model to evaluate potential sampling points in the input space, and determine the subsequent sampling strategy based on the evaluation results through the multi-fidelity predictive acquisition function; Step S3: Re-collect data using the determined sampling strategy, augment the re-collected data points into the dataset, and retrain the surrogate model using the dataset to achieve continuous adaptive learning that matches the global accuracy of the model, thus obtaining an optimized surrogate model. Step S4: Repeat steps S2 to S3, evaluate the model's compliance based on the performance evaluation indicators established in step S1, and finally obtain a proxy model with satisfactory accuracy.
2. The method for dynamic modeling of optical communication system samples according to claim 1, characterized in that, The initialization data is sampled using random sampling or Latin hypercube sampling.
3. The method for dynamic modeling of optical communication system samples according to claim 1, characterized in that, The proxy model is a serial co-kickinic proxy model, expressed as: in, It is any combination of parameters in the input space, corresponding to the adjustable variables of the optical communication simulation system; It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link; It is a low-fidelity function that provides an estimate of a given performance metric, based on a physical model; scaling factor measure right The global linear contribution is determined through the optimization process; the residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent of each other.
4. The method for dynamic modeling of optical communication system samples according to claim 3, characterized in that, In step S1, the modeling process of the proxy model is as follows: Collect sample points to obtain simulation data and experimental data ; The two components of the proxy model are modeled using independent Gaussian processes, and modeled using simulation data. Modeling using the difference between experimental and simulation data ; Using the complete dataset All hyperparameters in the surrogate model are jointly optimized using maximum likelihood estimation; hyperparameters include: kernel function parameters of the Gaussian process, scaling factor, etc. ; The trained proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance .
5. The method for dynamic modeling of optical communication system samples according to claim 1, characterized in that, In step S2, the multi-fidelity prediction acquisition function is used to balance information gain and sampling cost, and its expression is as follows: in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance; The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.
6. A dynamic modeling system for optical communication system samples, characterized in that, Includes the following modules: Module M1: Build an optical communication simulation system and corresponding experimental environment; collect simulation data and experimental data from the optical communication simulation system and experimental environment; initialize the proxy model using the simulation data and experimental data; establish performance evaluation indicators to evaluate the validity of the data and the model's compliance. Module M2: Establishes a multi-fidelity predictive acquisition function, uses a surrogate model to evaluate potential sampling points in the input space, and determines subsequent sampling strategies based on the evaluation results through the multi-fidelity predictive acquisition function; Module M3: Data is re-collected using a defined sampling strategy, the re-collected data points are augmented into the dataset, and the surrogate model is retrained using the dataset to achieve continuous adaptive learning that matches the global accuracy of the model, resulting in an optimized surrogate model. Module M4: Repeatedly executes modules M2 to M3, evaluates the model's compliance based on the performance evaluation indicators established in module M1, and finally obtains a proxy model that meets the accuracy requirements.
7. The dynamic modeling system for optical communication system samples according to claim 6, characterized in that, The initialization data is sampled using random sampling or Latin hypercube sampling.
8. The dynamic modeling system for optical communication system samples according to claim 6, characterized in that, The proxy model is a serial co-kickinic proxy model, expressed as: in, It is any combination of parameters in the input space, corresponding to the adjustable variables of the optical communication simulation system; It is a high-fidelity function, serving as the overall objective function of the surrogate model, representing the true performance of the fiber optic link; It is a low-fidelity function that provides an estimate of a given performance metric, based on a physical model; scaling factor measure right The global linear contribution is determined through the optimization process; the residual function Represents the scaled-up version Prediction and Reality Non-linear differences between values, and They are independent of each other.
9. The dynamic modeling system for optical communication system samples according to claim 8, characterized in that, In module M1, the modeling process of the proxy model is as follows: Collect sample points to obtain simulation data and experimental data ; The two components of the proxy model are modeled using independent Gaussian processes, and modeled using simulation data. Modeling using the difference between experimental and simulation data ; Using the complete dataset All hyperparameters in the surrogate model are jointly optimized using maximum likelihood estimation; hyperparameters include: kernel function parameters of the Gaussian process, scaling factor, etc. ; The trained proxy model can provide any input point. place Complete posterior probability distribution, predicted mean and variance .
10. The dynamic modeling system for optical communication system samples according to claim 6, characterized in that, In module M2, the multi-fidelity prediction acquisition function is used to balance information gain and sampling cost, and its expression is as follows: in, Calculate at point Data capture fidelity After the data, high-fidelity function The expected reduction in global prediction variance; The cost of data collection at different fidelity levels is represented by a weighted function in real-world scenarios.