Optical network rule discovery method and system based on data enhancement and decision tree

Abstract

The invention provides an optical network rule discovery method and system based on data enhancement and a decision tree, and the method comprises the steps: initializing a digital twin model, building a digital twin system based on an optical fiber network model and monitoring data, generating a simulation data set containing random configuration and a predicted transmission quality QoT value, training a generative model based on a generative adversarial network (GAN) by using the simulation data set; extracting high-performance samples with QoT values higher than a preset threshold in the real optical network data set, finely tuning the GAN-based generative model by using the high-performance samples, generating more high-performance samples by using the generative model, and splicing the generated high-performance samples with the original data set to form an enhanced data set; learning an interpretable decision tree classifier by taking the enhanced data set as a training set and taking Gini impurity as a division criterion, and disturbing and finely adjusting a judgment threshold value of each node of the decision tree; and extracting a high-performance optical network configuration rule based on the decision tree classifier.

Description

Technical Field

[0001] This invention relates to the field of optical communication network technology, and more specifically, to a method and system for optical network rule discovery based on data augmentation and decision trees. Background Technology

[0002] Real-world optical communication networks are complex, requiring the extraction of interpretable rules from network QoT data to guide network optimization. However, randomly collected datasets contain few high-performance data samples that yield high QoT results, impacting the effectiveness and accuracy of rule extraction.

[0003] The existing meta-learning-based small dataset learning method (published in the May 2022 issue of the Journal of Lightwave Technology, pp. 2684-2695, entitled "A Meta-Learning-Assisted Training Framework for Physical Layer Modeling in Optical Networks") can support data modeling and learning with a small number of data samples, but due to the black-box nature of neural networks, it is impossible to summarize interpretable rules and thus cannot reliably guide subsequent network optimization.

[0004] Patent application CN119362429A discloses a method and apparatus for predicting the power generation of a photovoltaic system based on edge clusters, applicable to distributed photovoltaic power generation systems. The method includes: acquiring, at each first preset time interval, first power generation, first location information, second power generation, second location information, first maintenance plan, and first environmental data, and preprocessing them to obtain a first training dataset; performing feature filtering on the first training dataset using a recursive feature elimination method to obtain a second training dataset; extracting features from the second training dataset using factor analysis to obtain a third training dataset; training a prediction model based on the third training dataset to obtain a power generation prediction model; obtaining target prediction parameters, inputting the target prediction parameters into the power generation prediction model, and obtaining the target power generation. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of this invention. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for optical network rule discovery based on data augmentation and decision trees.

[0006] The optical network rule discovery method based on data augmentation and decision trees provided by the present invention includes: Step 1: Initialize the digital twin model. Establish a digital twin system based on the fiber optic network model and monitoring data. Use the digital twin system to generate a simulation dataset containing random configuration and predicted transmission quality QoT values. Use the simulation dataset to train a generative model based on Generative Adversarial Network (GAN). Step 2: Extract high-performance samples with QoT values ​​higher than a preset threshold from the real optical network dataset, fine-tune the GAN-based generative model using the high-performance samples, generate more high-performance samples using the generative model, and concatenate the generated high-performance samples with the original dataset to form an enhanced dataset. Step 3: Using the augmented dataset as the training set, learn an interpretable decision tree classifier using Gini impurity as the splitting criterion, and perturb and fine-tune the decision thresholds of each node of the decision tree to improve the classification accuracy of high-performance samples. Step 4: Extract rules for high-performance optical network configuration based on decision tree classifier.

[0007] Preferably, in step 1, the digital twin model is based on the input optical network configuration vector. The expression for predicting the corresponding QoT metric vector Q is: ,in For the mapping function of the digital twin model; Training a generative model based on a Generative Adversarial Network (GAN) includes a generator and a discriminator. The generator takes random noise as input and outputs generated fake data samples. The discriminator takes samples as input and outputs the confidence scores of real samples. The training process includes alternating cycles of updating the discriminator and updating the generator. When updating the discriminator, real data samples are sampled from the simulation dataset and fake data samples are sampled from the generator. The discriminant loss is calculated and the discriminator parameters are updated. When updating the generator, noise is sampled as input to the generator to generate fake data samples. These samples are then input to the discriminator to obtain the output. The generator loss is constructed and the generator parameters are updated.

[0008] Preferably, in step 2, high-performance samples with QoT values ​​higher than a preset threshold are extracted from the real optical network dataset; the GAN-based generative model is fine-tuned using high-performance samples, including updating the discriminator and generator of the GAN with a small learning rate; more high-performance samples are generated using the generative model, including sampling random noise into the generator output new data samples, and selecting the portion with QoT values ​​higher than the preset threshold as the generated high-performance samples. The generated high-performance samples are concatenated with the original dataset to form an augmented dataset, in which the number of high-performance samples is equal to the amount of sample data in the original dataset.

[0009] Preferably, in step 3, an interpretable decision tree is learned using an augmented dataset as the training set and Gini impurity as the splitting criterion. The target variable is a binary classification label, with high-performance samples labeled as 1 and low-performance samples labeled as 0. The learning process includes traversing candidate features and their candidate thresholds for each node to be split, calculating the weighted Gini impurity of the left and right child nodes after splitting, selecting the feature-threshold combination that maximizes the reduction of the overall Gini impurity for splitting, and recursively splitting until the preset upper limit of tree depth is reached. The decision thresholds for each node of the decision tree are perturbed and fine-tuned. This includes applying small random perturbations to the thresholds of each split node, evaluating the accuracy of the model on high-performance samples after perturbation, repeating the process multiple times and recording the corresponding metrics, and selecting the threshold with the highest accuracy metric for each node to obtain the fine-tuned decision tree.

[0010] Preferably, in step 4, the rules for configuring high-performance optical networks are extracted based on the decision tree classifier, including extracting feature conditions from the paths that correctly classify high-performance samples and organizing them into an interpretable set of configuration rules. The characteristic conditions include optical network configuration parameters, which include input power and gain settings.

[0011] The optical network rule discovery system based on data augmentation and decision trees provided by the present invention includes: Module M1: Initialize the digital twin model, establish a digital twin system based on the fiber optic network model and monitoring data, use the digital twin system to generate a simulation dataset containing random configuration and predicted transmission quality QoT values, and use the simulation dataset to train a generative model based on generative adversarial network (GAN). Module M2: Extract high-performance samples with QoT values ​​higher than a preset threshold from real optical network datasets, fine-tune the GAN-based generative model using high-performance samples, generate more high-performance samples using the generative model, and concatenate the generated high-performance samples with the original dataset to form an enhanced dataset. Module M3: Using the augmented dataset as the training set, it learns an interpretable decision tree classifier using Gini impurity as the splitting criterion, and perturbs and fine-tunes the decision thresholds of each node of the decision tree to improve the classification accuracy of high-performance samples. Module M4: Extracts rules for high-performance optical network configuration based on a decision tree classifier.

[0012] Preferably, in module M1, the digital twin model is based on the input optical network configuration vector. The expression for predicting the corresponding QoT metric vector Q is: ,in For the mapping function of the digital twin model; Training a generative model based on a Generative Adversarial Network (GAN) includes a generator and a discriminator. The generator takes random noise as input and outputs generated fake data samples. The discriminator takes samples as input and outputs the confidence scores of real samples. The training process includes alternating cycles of updating the discriminator and updating the generator. When updating the discriminator, real data samples are sampled from the simulation dataset and fake data samples are sampled from the generator. The discriminant loss is calculated and the discriminator parameters are updated. When updating the generator, noise is sampled as input to the generator to generate fake data samples. These samples are then input to the discriminator to obtain the output. The generator loss is constructed and the generator parameters are updated.

[0013] Preferably, in module M2, high-performance samples with QoT values ​​higher than a preset threshold are extracted from the real optical network dataset; the GAN-based generative model is fine-tuned using high-performance samples, including updating the discriminator and generator of the GAN with a small learning rate; more high-performance samples are generated using the generative model, including sampling random noise into the generator output new data samples, and selecting the portion with QoT values ​​higher than the preset threshold as the generated high-performance samples. The generated high-performance samples are concatenated with the original dataset to form an augmented dataset, in which the number of high-performance samples is equal to the amount of sample data in the original dataset.

[0014] Preferably, in module M3, an interpretable decision tree is learned using an augmented dataset as the training set and Gini impurity as the splitting criterion. The target variable is a binary classification label, with high-performance samples labeled as 1 and low-performance samples labeled as 0. The learning process includes traversing candidate features and their candidate thresholds for each node to be split, calculating the weighted Gini impurity of the left and right child nodes after splitting, selecting the feature-threshold combination that maximizes the reduction of the overall Gini impurity for splitting, and recursively splitting until the preset upper limit of tree depth is reached. The decision thresholds for each node of the decision tree are perturbed and fine-tuned. This includes applying small random perturbations to the thresholds of each split node, evaluating the accuracy of the model on high-performance samples after perturbation, repeating the process multiple times and recording the corresponding metrics, and selecting the threshold with the highest accuracy metric for each node to obtain the fine-tuned decision tree.

[0015] Preferably, in module M4, the rules for extracting high-performance optical network configuration based on the decision tree classifier include extracting feature conditions from the paths that correctly classify high-performance samples and organizing them into an interpretable set of configuration rules. The characteristic conditions include optical network configuration parameters, which include input power and gain settings.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention solves the problem of difficulty in effectively extracting rules from real optical network datasets based on only a small number of high-performance samples by adopting a data mining structure that combines data augmentation and decision trees. It achieves the effect of obtaining interpretable, high-precision and executable configuration rules in optical networks. Attached Figure Description

[0017] 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 Flowchart of a data augmentation and decision tree-based optical network rule discovery method; Figure 2 For decision tree classifiers. Detailed Implementation

[0018] 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.

[0019] Example 1 This invention provides a method for optical network rule discovery based on data augmentation and decision trees. It enhances high-performance samples in an optical network dataset using a generative model to improve the diversity and coverage of high-quality configurations. Subsequently, a decision tree classifier is trained to extract rules for interpretable high-performance optical network configurations.

[0020] like Figure 1 The specific implementation steps are as follows: Step 1: Initialize the digital twin model and generate a simulation dataset to train the generative model; Step 1.1: Establish a digital twin system based on the fiber optic network model and monitoring data; Digital twin models can predict the corresponding transmission quality (QoT) based on the input optical network configuration, expressed as: ,in Represents the QoT metric vector. A mapping representing a digital twin model, This represents the configuration vector of the optical network.

[0021] Step 1.2: Use a digital twin system to generate a simulation dataset containing random configuration and predicted transmission quality (QoT) values; Step 1.3: Train a generative model based on Generative Adversarial Networks (GANs) using a simulation dataset.

[0022] Simulation Dataset It contains simulation data generated from multiple digital twin models, each Includes network configuration and its corresponding predicted QoT value ; GAN-based generative models include generators. and discriminator The generator is used to generate new samples; it takes random noise as input and can output generated fake data samples. The sample has the same shape as the real sample. The discriminator is used to determine whether the sample is real or fake. The input sample... It can output the confidence level that a sample is a true sample.

[0023] When updating the discriminator, sample some real data from the dataset. Simultaneously sample noise input generator Generate some fake data . The corresponding annotation is true. The corresponding label is false, input to the discriminator. Perform discrimination and calculate the loss function for discrimination. Only backpropagate and update the discriminator parameters to make them distinguish between real and fake samples as accurately as possible.

[0024] When updating the generator, sampled noise is input into the generator to generate fake data, which is then fed into the discriminator. Perform discrimination and obtain the output. Based on the discriminator... The output is used to construct the generator loss (the goal is to make the discriminator more inclined to judge it as true). Only the generator parameters are backpropagated and updated, so that the generated samples gradually "deceive" the discriminator during training.

[0025] Training alternates between updating the discriminator and the generator. After training converges, the generator... It can generate samples from noise that are difficult to distinguish from real data.

[0026] Step 2: Enhance the high-performance samples of the optical network dataset using generative models to generate an enhanced dataset; Step 2.1: Extract high-performance samples from real optical network datasets whose QoT values ​​are higher than a certain threshold; Step 2.2: Fine-tune the GAN-based generative model using high-performance samples, and then use the generative model to generate more high-performance samples. Similar to step 1.3, the discriminator and generator of the GAN are updated using high-performance samples from the real dataset. A small learning rate is used during the update, and the GAN is only fine-tuned so that it can learn the distribution of high-performance samples from the real dataset.

[0027] A generator that samples random noise as input to a generative model can output new data samples based on the learned distribution of high-performance samples. . Includes newly generated network configuration And the corresponding prediction performance. From the newly generated data samples, those with QoT values ​​higher than a certain threshold are selected as high-performance samples.

[0028] Step 2.3: Concatenate the generated high-performance samples with the original dataset to form an enhanced dataset.

[0029] The generated high-performance samples are added to the original dataset, so that the amount of data for high-performance samples and ordinary performance samples is roughly equal, forming an enhanced dataset.

[0030] Step 3: Train an interpretable decision tree classifier based on the augmented dataset; Step 3.1: Using the augmented dataset as the training set, and using Gini impurity as the splitting criterion, learn an interpretable decision tree to achieve the optimal split between high-performance and low-performance samples; During segmentation, the target variable is first defined as a binary classification label (high performance = 1, low performance = 0). During the training phase, for each node to be segmented, candidate features and their candidate thresholds are traversed, the weighted Gini impurity of the left and right child nodes after the segmentation is calculated, and the feature-threshold combination that maximizes the decrease in overall Gini impurity (i.e., maximizes the purity increase) is selected for segmentation; the same criteria are used for recursive segmentation until the preset upper limit of tree depth is reached.

[0031] Step 3.2: Perturb and fine-tune the decision thresholds of each node in the optimal splitting decision tree to further improve the classification accuracy of high-performance samples.

[0032] Since the optimization objective focuses on the discrimination accuracy of high-performance samples (label=1), simply using Gini impurity for splitting may not be optimal for the positive class. Therefore, based on the trained decision tree, a small random perturbation is applied to the threshold of each split node, and the accuracy of the model on high-performance samples after the perturbation is evaluated. This process is repeated multiple times for different perturbations, and the corresponding metrics are recorded. Finally, the threshold setting that maximizes the accuracy metric is selected for each node, resulting in a fine-tuned decision tree.

[0033] Step 4: Extract rules for high-performance optical network configuration based on decision tree classifier; Extract feature conditions from the path of correctly classifying high-performance samples, and organize them into an interpretable and executable set of configuration rules.

[0034] For example, in a certain training iteration, we used the input power of four fiber optic spans (input_power1~input_power4) and the gain settings of six amplifiers (gain1~gian6) as the splitting features of the decision tree, and obtained the following: Figure 2 The decision tree classifier shown is used (high-performance samples are class1, and other samples are class0).

[0035] As seen from the leaf nodes of the tree, there are two paths leading to class=1. To achieve a higher performance hit rate, we choose the path with the highest proportion of positive classes (the darkest blue path): Input_power3<=21.45 is true, Input_power4<=14.65 is false, and Input_power2<=11.15 is false.

[0036] Based on this, the following interpretable high-performance configuration rules can be summarized: The third segment input power (Input_power3) is recommended to be ≤21.45dB (to avoid being too high); the fourth segment input power (Input_power4) is recommended to be >14.65dB (to avoid being too low); and the second segment input power (Input_power2) is recommended to be >11.15dB (to avoid being too low). These rules can be further explained in conjunction with the segment lengths and attenuation parameters.

[0037] This rule is derived from the highest proportion of class=1 in the leaf nodes of the corresponding path, reflecting the most favorable configuration constraints for high-performance samples under the current data and model assumptions. The rule is interpretable and actionable, providing direct guidance for the practical optimization of optical networks.

[0038] Example 2 This invention also provides a rule discovery system for optical networks based on data augmentation and decision trees, comprising: Module M1: Initialize the digital twin model, establish a digital twin system based on the fiber optic network model and monitoring data, use the digital twin system to generate a simulation dataset containing random configuration and predicted transmission quality QoT values, and use the simulation dataset to train a generative model based on generative adversarial network (GAN). Module M2: Extract high-performance samples with QoT values ​​higher than a preset threshold from real optical network datasets, fine-tune the GAN-based generative model using high-performance samples, generate more high-performance samples using the generative model, and concatenate the generated high-performance samples with the original dataset to form an enhanced dataset. Module M3: Using the augmented dataset as the training set, it learns an interpretable decision tree classifier using Gini impurity as the splitting criterion, and perturbs and fine-tunes the decision thresholds of each node of the decision tree to improve the classification accuracy of high-performance samples. Module M4: Extracts rules for high-performance optical network configuration based on a decision tree classifier.

[0039] In module M1, the digital twin model is based on the input optical network configuration vector. The expression for predicting the corresponding QoT metric vector Q is: ,in For the mapping function of the digital twin model; Training a generative model based on a Generative Adversarial Network (GAN) includes a generator and a discriminator. The generator takes random noise as input and outputs generated fake data samples. The discriminator takes samples as input and outputs the confidence scores of real samples. The training process includes alternating cycles of updating the discriminator and updating the generator. When updating the discriminator, real data samples are sampled from the simulation dataset and fake data samples are sampled from the generator. The discriminant loss is calculated and the discriminator parameters are updated. When updating the generator, noise is sampled as input to the generator to generate fake data samples. These samples are then input to the discriminator to obtain the output. The generator loss is constructed and the generator parameters are updated.

[0040] In module M2, high-performance samples with QoT values ​​higher than a preset threshold are extracted from real optical network datasets; the GAN-based generative model is fine-tuned using high-performance samples, including updating the discriminator and generator of the GAN with a small learning rate; more high-performance samples are generated using the generative model, including sampling random noise into the generator to output new data samples, and selecting the portion with QoT values ​​higher than the preset threshold as the generated high-performance samples. The generated high-performance samples are concatenated with the original dataset to form an augmented dataset, in which the number of high-performance samples is equal to the amount of sample data in the original dataset.

[0041] In module M3, an augmented dataset is used as the training set, and Gini impurity is used as the splitting criterion to learn an interpretable decision tree. The target variable is a binary classification label, with high-performance samples labeled as 1 and low-performance samples labeled as 0. The learning process includes traversing candidate features and their candidate thresholds for each node to be split, calculating the weighted Gini impurity of the left and right child nodes after splitting, selecting the feature-threshold combination that maximizes the reduction of the overall Gini impurity for splitting, and recursively splitting until the preset tree depth limit is reached. The decision thresholds for each node of the decision tree are perturbed and fine-tuned. This includes applying small random perturbations to the thresholds of each split node, evaluating the accuracy of the model on high-performance samples after perturbation, repeating the process multiple times and recording the corresponding metrics, and selecting the threshold with the highest accuracy metric for each node to obtain the fine-tuned decision tree.

[0042] In module M4, rules for configuring high-performance optical networks are extracted based on a decision tree classifier, including extracting feature conditions from paths that correctly classify high-performance samples and organizing them into an interpretable set of configuration rules. The characteristic conditions include optical network configuration parameters, which include input power and gain settings.

[0043] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0044] 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 optical network rule discovery based on data augmentation and decision tree, characterized in that, The method comprises the following steps: Step 1: initializing a digital twin model, establishing a digital twin system based on a fiber network model and monitoring data, generating a simulation data set containing random configurations and predicted transmission quality QoT values by using the digital twin system, and training a generative model based on a generative adversarial network (GAN) by using the simulation data set; Step 2: extracting high-performance samples with QoT values higher than a preset threshold from a real optical network data set, fine-tuning the generative model based on GAN by using the high-performance samples, generating more high-performance samples by using the generative model, and splicing the generated high-performance samples and the original data set to form an enhanced data set; Step 3: using the enhanced data set as a training set, learning an interpretable decision tree classifier by using Gini impurity as a division criterion, perturbing and fine-tuning the threshold values of nodes of the decision tree to improve the classification accuracy of the high-performance samples; Step 4: extracting rules of high-performance optical network configurations based on the decision tree classifier.

2. The data augmentation and decision tree based optical network rule discovery method of claim 1, wherein, In the step 1, the digital twin model is based on the input optical network configuration vector The corresponding QoT index vector Q is predicted, expressed as: Wherein is the mapping function of the digital twin model; The training of the generative model based on the generative adversarial network (GAN) comprises a generator and a discriminator, wherein the generator inputs random noise and outputs generated fake data samples; the discriminator inputs samples and outputs the confidence of the samples being true samples; the training process comprises an alternating cycle of updating the discriminator and updating the generator, wherein when the discriminator is updated, true data samples are sampled from the simulation data set and fake data samples are sampled from the generator, the discrimination loss is calculated and the discriminator parameters are updated; when the generator is updated, noise is sampled to input the generator to generate fake data samples, the output is obtained by inputting the discriminator, the generator loss is constructed and the generator parameters are updated.

3. The data augmentation and decision tree based optical network rule discovery method of claim 1, wherein, In step 2, high-performance samples with QoT values higher than a preset threshold are extracted from a real optical network data set; the generative model based on GAN is fine-tuned by using the high-performance samples, which comprises updating the discriminator and the generator of GAN using a small learning rate; more high-performance samples are generated by using the generative model, which comprises sampling random noise to input the generator to output new data samples, and selecting the part with QoT values higher than the preset threshold as the generated high-performance samples; The generated high-performance samples are spliced with the original data set to form an enhanced data set, wherein the number of high-performance samples is the same as the amount of sample data in the original data set.

4. The data augmentation and decision tree based optical network rule discovery method of claim 1, wherein, In step 3, the enhanced data set is used as a training set, and an interpretable decision tree is learned by using Gini impurity as a division criterion, wherein the target variable is a binary classification label, the label of the high-performance sample is 1, and the label of the non-high-performance sample is 0; the learning process comprises traversing candidate features and their candidate threshold values for each node to be divided, calculating the weighted Gini impurity of the left and right child nodes after division, selecting the feature-threshold combination that makes the total Gini impurity decrease most to perform division, and recursively dividing until a preset upper limit of tree depth is reached; The threshold values of nodes of the decision tree are perturbed and fine-tuned, which comprises applying small random perturbations to the threshold values of each split node, evaluating the accuracy of the model on high-performance samples after perturbation, repeating multiple times and recording the corresponding indicators, selecting the threshold value setting with the highest accuracy rate indicator for each node, and obtaining the fine-tuned decision tree.

5. The data augmentation and decision tree based optical network rule discovery method of claim 1, wherein, In step 4, rules for extracting high-performance optical network configurations based on the decision tree classifier include extracting feature conditions from paths that correctly classify high-performance samples and organizing them into a set of interpretable configuration rules. The feature conditions include optical network configuration parameters, including input power and gain settings.

6. An optical network rule discovery system based on data augmentation and decision trees, characterized in that, It includes: Module M1: initialize the digital twin model, establish a digital twin system based on the fiber network model and monitoring data, use the digital twin system to generate a simulation data set containing random configurations and predicted transmission quality QoT values, and train a generative model based on a generative adversarial network (GAN) using the simulation data set; Module M2: extract high-performance samples with QoT values higher than a preset threshold from the real optical network data set, fine-tune the GAN-based generative model using the high-performance samples, generate more high-performance samples using the generative model, and concatenate the generated high-performance samples with the original data set to form an enhanced data set; Module M3: use the enhanced data set as the training set, use Gini impurity as the division criterion to learn an interpretable decision tree classifier, and perturb and fine-tune the threshold values of each node of the decision tree to improve the classification accuracy of high-performance samples; Module M4: extract rules for high-performance optical network configurations based on the decision tree classifier.

7. The data augmentation and decision tree based optical network rule discovery system of claim 6, wherein, In said module M1, said digital twin model is based on an input optical network configuration vector a corresponding QoT indicator vector Q is predicted, expressed as: wherein is a mapping function of the digital twin model; Training a generative model based on a generative adversarial network (GAN) includes a generator and a discriminator, where the generator inputs random noise and outputs generated fake data samples; the discriminator inputs samples and outputs the confidence that the sample is a true sample; the training process includes an alternating cycle of updating the discriminator and updating the generator, where when updating the discriminator, true data samples are sampled from the simulation data set and fake data samples are sampled from the generator, the discrimination loss is calculated and the discriminator parameters are updated; when updating the generator, sample noise is input into the generator to generate fake data samples, input into the discriminator to get the output, construct the generator loss and update the generator parameters.

8. The data augmentation and decision tree based optical network rule discovery system of claim 6, wherein, In module M2, high-performance samples with QoT values higher than a preset threshold are extracted from the real optical network data set; the GAN-based generative model is fine-tuned using the high-performance samples, including updating the discriminator and generator of the GAN using a small learning rate; more high-performance samples are generated using the generative model, including sampling random noise to input into the generator to output new data samples, and selecting the part with QoT values higher than the preset threshold as the generated high-performance samples; The generated high-performance samples are concatenated with the original data set to form an enhanced data set, where the number of high-performance samples is equal to the amount of sample data in the original data set.

9. The data augmentation and decision tree based optical network rule discovery system of claim 6, wherein, In module M3, the enhanced data set is used as the training set, and Gini impurity is used as the division criterion to learn an interpretable decision tree, where the target variable is a binary classification label, the high-performance sample label is 1, and the non-high-performance sample label is 0; the learning process includes traversing candidate features and their candidate thresholds for each node to be divided, calculating the weighted Gini impurity of the left and right child nodes after division, selecting the feature-threshold combination that maximizes the total Gini impurity decrease for cutting, and recursively dividing until the preset upper limit of tree depth is reached; The decision tree threshold of each node is disturbed and fine-tuned, including applying small random disturbance to the threshold of each split node, evaluating the accuracy of the disturbed model on high-performance samples, repeating multiple times and recording the corresponding indicators, selecting the threshold setting with the highest accuracy indicator for each node, and obtaining the fine-tuned decision tree.

10. The data augmentation and decision tree based optical network rule discovery system of claim 6, wherein, In the module M4, rules for extracting high-performance optical network configurations based on decision tree classifiers are extracted, including extracting feature conditions from paths that correctly classify high-performance samples, and organizing them into an interpretable configuration rule set. The feature conditions include optical network configuration parameters, and the optical network configuration parameters include input power and gain settings.

Images