A method, device, equipment and medium for generating a simulation topology of a telecommunication transmission network

Abstract

The present disclosure relates to the technical field of telecommunications transmission, and particularly relates to a method and device for generating a simulation topology of a telecommunications transmission network, and a medium. The present disclosure deeply integrates knowledge rules in the telecommunications field and generative GAN, and solves the problem that fidelity and compliance cannot be considered simultaneously in the generation of a simulation topology. The present disclosure can simultaneously generate a highly realistic network structure, static configuration and dynamic traffic. Through hard rule constraints, it is ensured that the generated result conforms to the design specifications of a telecommunications network, such as core double nodes and access double homing. End-to-end one-key generation from data to a directly deployable simulation environment is realized, and the efficiency is greatly improved. The output of the present disclosure is a standard network management model, which is seamlessly connected to a live network management and control system through a standard protocol, and is ready for use out of the box.

Description

Technical Field

[0001] This disclosure relates to the field of telecommunications transmission technology, and in particular to a method, apparatus, equipment and medium for generating a simulation topology of a telecommunications transmission network. Background Technology

[0002] Modern telecommunications transmission networks, such as OTN, PTN, and IP-RAN, require centralized management and control systems for service activation, fault diagnosis, and resource scheduling. During the development and testing phase, a highly realistic network topology environment must be constructed to verify the functionality (such as path calculation and fault switching) and performance (such as large-scale network element management capabilities) of the management and control system.

[0003] Real-world telecommunications transmission network topologies exhibit significant characteristics, including a structured hierarchy that follows a "core layer-aggregation layer-access layer" logical architecture and is strongly correlated with geographical administrative levels (province-city-district / county). They also possess parameter complexity, such as static parameters (e.g., network element type (router / OTN device), port density, optical module power) and dynamic parameters (e.g., traffic spatiotemporal fluctuations, link latency jitter, protection group status). Simulated topologies must closely approximate real networks in both structural and parameter dimensions; otherwise, the test results of the management and control system will be distorted.

[0004] In the existing technology, there are several methods for generating simulation topology: manual specification, where engineers draw the topology based on experience; template reuse, where preset templates (such as ring and star topologies) are called; and random generation, where network elements are randomly connected according to a probability model.

[0005] However, existing technologies suffer from topology and parameter distortion. Topology distortion includes: manual specification is limited by personal experience and cannot cover the topological differences across multiple regions, such as sparse mountain networks versus high-density urban networks, resulting in a single topology that cannot represent the diversity of the entire network; the template library for template reuse is limited in size and cannot adapt to complex hybrid architectures, such as nested ring networks and mesh networks, resulting in high topology repetition and a lack of the irregular characteristics of real networks; random generation ignores network construction rules, such as high interconnectivity of core nodes and single aggregation of access nodes, leading to a messy topology; and it generates "reverse engineering" structures, such as isolated nodes and redundant rings. Parameter distortion includes: static parameter distortion, such as independent and random allocation of parameters like device model and port quantity, which does not follow the statistical laws of real networks, such as core nodes having more than 80% 100G ports, indicating incorrect resource allocation logic and distorted capacity planning tests of the management system; and dynamic parameter distortion, such as the inability to simulate the spatiotemporal characteristics of real business traffic, such as peak-hour fluctuations and sudden traffic surges during holidays, making key performance verifications such as fault recovery latency and congestion control unreliable.

[0006] Existing methods lack the ability to quantitatively model the engineering patterns of telecommunications networks, resulting in simulation topologies that, at the structural level, fail to meet the conflicting requirements of "regional diversity" and "engineering regularity"; at the parameter level, static parameters are discrete and distorted, and dynamic parameters lack temporal sequence. This leads to the control system passing tests in low-fidelity topologies but exhibiting functional defects and performance bottlenecks in real networks. Summary of the Invention

[0007] To address the aforementioned issues, this disclosure provides a method, apparatus, device, and medium for generating simulation topology for telecommunications transmission networks, which solves problems related to topology generation, parameter system generation, control system adaptation, and simulation efficiency.

[0008] A first aspect includes a method for generating a simulated topology for a telecommunications transmission network, the method comprising: Obtain engineering topology datasets of telecommunications transmission networks, extract multidimensional telecommunications features, construct conditional vectors, and generate standardized spatiotemporal graph data representations; Based on standardized spatiotemporal graph data representation, a generator for generative adversarial networks (GANs) is constructed, and a discriminator for the GANs is constructed using a multi-scale evaluation branch. Establish a loss function and train the generative adversarial network to obtain the trained network; Based on the engineering topology dataset of real-time telecommunications transmission networks, topology graph structure data is generated through trained networks; The generated topology data is converted into a network management model to obtain the network management configuration file.

[0009] Furthermore, the engineering topology dataset for telecommunications transmission networks includes: Configuration backup library or performance data collection library for the existing network management system and network element management system.

[0010] Furthermore, an engineering topology dataset of the telecommunications transmission network is obtained, and multidimensional telecommunications features are extracted, including: After data acquisition, data cleaning and alignment are performed, including handling missing values, standardizing device naming conventions, and aligning data timestamps from different system sources to form a consistent global network view.

[0011] Furthermore, multi-dimensional telecommunications features are extracted, including: network element node attributes, link attributes, time-series data of service traffic matrix, and historical fault event data; Network element node attributes include: device type, geographical level, network level, device model, rack slot capacity, number of used ports and port speed, and device weight; Link attributes include: physical fiber type, link length, logical link bandwidth, link cost, protection group identifier, and optical power attenuation value; The service traffic matrix time series data includes: network element-level inbound and outbound traffic, critical link utilization, and bandwidth time series data of end-to-end service flows collected at fixed time granularity; Historical fault event data includes: event ID, timestamp, triggering network element and port, fault type, associated protection group, switching status, and switching duration.

[0012] Furthermore, construct the condition vector, including: For each network element node in the dataset, the category label is extracted, and one-hot encoding is performed on the category label. The category label of each network element node is converted into a high-dimensional one-hot vector. The category label includes: geographic level label and network level label. The high-dimensional one-hot vectors of different category labels are concatenated along the feature dimension to form a sparse vector; the sparse vector is then mapped to a low-dimensional dense continuous vector space through an embedding layer, and the initial condition vector of the node is output. The initial condition vector is used as a trainable parameter. During the entire training process of the generative adversarial network, the final condition vector is obtained through joint optimization using the backpropagation algorithm and gradient descent.

[0013] Furthermore, the embedding layer employs a multilayer perceptron mapping, inputting sparse vectors into the multilayer perceptron (MLP) and outputting low-dimensional dense vectors as initial condition vectors; wherein, the MLP includes: a linear layer and a nonlinear activation function.

[0014] Furthermore, a standardized spatiotemporal graph data representation is generated, including: Generate a standardized spatiotemporal graph data representation: G = (A, X, T), where: A represents the adjacency matrix; X represents the node feature matrix; T represents the dynamic parameter time sequence matrix.

[0015] Furthermore, the generator for constructing generative adversarial networks includes: The generator's input is a noise vector Z plus a condition vector C; where the noise vector Z and the condition vector C are concatenated at the input layer. The generator outputs topology graph structure data G'= (A', X', T') constrained by telecommunications rules; Among them, telecommunications rule constraints include: full interconnection constraints of core layer nodes and dual aggregation constraints of access nodes; A' represents the adjacency matrix; X' represents the node feature matrix; T' represents the dynamic parameter time series matrix; The constraint for full interconnection of core layer nodes is that A' core node degree ≥ threshold; the constraint for dual aggregation of access nodes is that A' access node degree = set value.

[0016] Furthermore, a discriminator for the generative adversarial network is constructed using a multi-scale evaluation branch, including: The discriminator includes: topology branch, static attribute branch, and dynamic behavior branch; The topology branch is based on the graph convolutional network GCN or the graph attention network GAT. It receives standardized spatiotemporal graph data representation or topology graph structure data, aggregates neighbor information through message passing mechanism, and finally outputs graph-level embedding vectors. The embedding vectors are used to determine whether the connection mode and hierarchical structure of the entire graph are consistent with the real telecommunications network. The static attribute branch is a multilayer perceptron (MLP); it receives the statistical features or histogram of the node feature matrix X; the output is used to calculate the JS divergence or Wasserstein distance between the data and the real data distribution. The dynamic behavior branch analyzes node traffic sequences using a temporal neural network; the temporal neural network extracts the time and frequency domain features of the traffic sequences by integrating Fast Fourier Transform (FFT), and compares the differences between generated traffic and real traffic in terms of periodicity, fluctuation patterns, and suddenness. The output feature vectors of the topology branch, static attribute branch, and dynamic behavior branch are concatenated and then aggregated through a fully connected layer to generate the final authenticity score.

[0017] Furthermore, the loss function is established, including: The generator's total loss function L_G is a weighted sum of adversarial loss, node degree distribution loss, and device distribution loss; The generator's total loss function L_G is given by the following formula: L_G=E[log(1 D(G'))]+λ1 ||Degree(A') - Degree_real||2+λ2 JS(P_model ||P_real); Where D(G): The discriminator's evaluation output of the real topology. It represents the probability or score that the discriminator network classifies the real topology data G as "true" (in Wasserstein GAN, it is an unconstrained score). The higher this value, the more credible the discriminator considers the real data.

[0018] D(G'): The discriminator's evaluation output of the generated topology. It represents the probability or score by which the discriminator network classifies the simulated topology G' generated by the generator as "true". The higher the value, the more deceptive the generated topology is and the higher its quality. The generator's goal is to maximize E[D(G')].

[0019] λ1, λ2, λ3: Weighting hyperparameters of the loss function. These are real numbers greater than 0, used to balance the importance and dimensions of different component losses in the total loss function. For example: λ1 controls the weight of the node degree distribution loss.

[0020] λ2 controls the weights of the distribution loss (JS divergence) of the control device type.

[0021] λ3 controls the weight of the dynamic time loss (DTW distance).

[0022] The specific values ​​of these hyperparameters are usually determined through grid search or empirical tuning.

[0023] Degree(A'): The sequence of node degrees that generate the topology. This is a vector, not a single numerical value. Each element represents the number of connections (degrees) of the corresponding node in the adjacency matrix A' of the generated topology. Degree(A') = [deg'(v1), deg'(v2), ..., deg'(v_N)].

[0024] Degree_real: The sequence of node degrees in the real topology. It is also a vector, where each element represents the degree of the corresponding node in the adjacency matrix A of the real topology. Degree_real = [deg_real(v1), deg_real(v2), ...,deg_real(v_N)].

[0025] || ... ||2: L2 norm (Euclidean norm). In this context, it is used to calculate the difference between two degree sequence vectors, Degree(A') and Degree_real. It is calculated as the square root of the sum of the squares of the differences between corresponding elements of the two vectors, and can comprehensively measure the overall deviation of the two degree distributions.

[0026] JS: Jensen-Shannon divergence. A symmetric index used to measure the similarity between two probability distributions. Its value range is [0, 1], with smaller values ​​indicating closer similarities between the two distributions. In this invention, it is used to compare the consistency between the generated topology and the actual topology in terms of device type distribution.

[0027] P_model: Empirical distribution of device types in the generated topology. This is a discrete probability distribution describing the frequency of occurrence of various types of devices (such as OTN, PTN) in the generated topology.

[0028] P_real: Empirical distribution of device types in the real topology. It is also a discrete probability distribution that describes the frequency of occurrence of various types of devices in a real network.

[0029] The adversarial loss E[log(1 - D(G'))] is used to induce the generation of a topology deception discriminator; Node degree distribution loss λ1 ||Degree(A') - Degree_real||2, used to force the generated topology to have a node degree distribution similar to the real network; Equipment distributed loss λ2 JS(P_model || P_real) represents minimizing the JS divergence between the generated topology and the real topology in terms of device type distribution; The discriminator's loss function is the standard adversarial loss L_D, as shown in the following formula: L_D = -[E[log(D(G))] + E[log(1 - D(G'))]]; Where E[log(D(G))] is the expected probability of the real data being judged as true; E[log(1 - D(G'))] is the expected probability of the generated data being judged as false.

[0030] Furthermore, the generated topology data is converted into a network management model, including: According to the mapping rules, the topology diagram data is converted into management objects defined by the telecommunications management network or the YANG model; the mapping rules include node mapping and link mapping. Generate network management configuration files in XML or JSON format that can be recognized by the network management system.

[0031] Furthermore, node mapping includes mapping features in the feature matrix of topology graph structure data nodes to device attributes; Link mapping includes converting the connection relationships in the adjacency matrix of the topology graph structure into connection lines or cross connections between ports, and organizing them into a containment tree structure that describes parent-child relationships.

[0032] Furthermore, it also includes: Establish a session with the network management interface of the target simulation environment via the Netconf or TL1 protocol, specify the target configuration database using the network management configuration file, and initiate a transactional configuration push request. After receiving the configuration, the target simulation environment automatically creates the corresponding virtual devices and establishes connections, constructing a simulation network consistent with the topology data, which is used for operation and maintenance drills, algorithm testing, or capacity planning.

[0033] Secondly, a telecommunications transmission network simulation topology generation device includes: Data construction unit, adversarial network construction unit, training unit, topology generation unit, and data transformation unit; The data construction unit is used to acquire engineering topology datasets of telecommunications transmission networks, extract multidimensional telecommunications features, construct conditional vectors, and generate standardized spatiotemporal graph data representations. The adversarial network building unit, based on standardized spatiotemporal graph data representation, is used to build the generator of the generative adversarial network, and a multi-scale evaluation branch is used to build the discriminator of the generative adversarial network. The training unit is used to establish the loss function and train the generative adversarial network to obtain the trained network. The topology generation unit is used to generate topology graph structure data from a trained network based on a real-time telecommunications transmission network engineering topology dataset. The data conversion unit is used to convert the generated topology diagram data into a network management model to obtain a network management configuration file.

[0034] Furthermore, the engineering topology dataset for telecommunications transmission networks includes: Configuration backup library or performance data collection library for the existing network management system and network element management system.

[0035] Furthermore, an engineering topology dataset of the telecommunications transmission network is obtained, and multidimensional telecommunications features are extracted, including: After data acquisition, data cleaning and alignment are performed, including handling missing values, standardizing device naming conventions, and aligning data timestamps from different system sources to form a consistent global network view.

[0036] Furthermore, multi-dimensional telecommunications features are extracted, including: network element node attributes, link attributes, time-series data of service traffic matrix, and historical fault event data; Network element node attributes include: device type, geographical level, network level, device model, rack slot capacity, number of used ports and port speed, and device weight; Link attributes include: physical fiber type, link length, logical link bandwidth, link cost, protection group identifier, and optical power attenuation value; The service traffic matrix time series data includes: network element-level inbound and outbound traffic, critical link utilization, and bandwidth time series data of end-to-end service flows collected at fixed time granularity; Historical fault event data includes: event ID, timestamp, triggering network element and port, fault type, associated protection group, switching status, and switching duration.

[0037] Furthermore, construct the condition vector, including: For each network element node in the dataset, the category label is extracted, and one-hot encoding is performed on the category label. The category label of each network element node is converted into a high-dimensional one-hot vector. The category label includes: geographic level label and network level label. The high-dimensional one-hot vectors of different category labels are concatenated along the feature dimension to form a sparse vector; the sparse vector is then mapped to a low-dimensional dense continuous vector space through an embedding layer, and the initial condition vector of the node is output. The initial condition vector is used as a trainable parameter. During the entire training process of the generative adversarial network, the final condition vector is obtained through joint optimization using the backpropagation algorithm and gradient descent.

[0038] Furthermore, the embedding layer employs a multilayer perceptron mapping, inputting sparse vectors into the multilayer perceptron (MLP) and outputting low-dimensional dense vectors as initial condition vectors; wherein, the MLP includes: a linear layer and a nonlinear activation function.

[0039] Furthermore, a standardized spatiotemporal graph data representation is generated, including: Generate a standardized spatiotemporal graph data representation: G = (A, X, T), where: A represents the adjacency matrix; X represents the node feature matrix; T represents the dynamic parameter time sequence matrix.

[0040] Furthermore, the generator for constructing generative adversarial networks includes: The generator's input is a noise vector Z plus a condition vector C; where the noise vector Z and the condition vector C are concatenated at the input layer. The generator outputs topology graph structure data G'= (A', X', T') constrained by telecommunications rules; Among them, telecommunications rule constraints include: full interconnection constraints of core layer nodes and dual aggregation constraints of access nodes; A' represents the adjacency matrix; X' represents the node feature matrix; T' represents the dynamic parameter time series matrix; The constraint for full interconnection of core layer nodes is that A' core node degree ≥ threshold; the constraint for dual aggregation of access nodes is that A' access node degree = set value.

[0041] Furthermore, a discriminator for the generative adversarial network is constructed using a multi-scale evaluation branch, including: The discriminator includes: topology branch, static attribute branch, and dynamic behavior branch; The topology branch is based on the graph convolutional network GCN or the graph attention network GAT. It receives standardized spatiotemporal graph data representation or topology graph structure data, aggregates neighbor information through message passing mechanism, and finally outputs graph-level embedding vectors. The embedding vectors are used to determine whether the connection mode and hierarchical structure of the entire graph are consistent with the real telecommunications network. The static attribute branch is a multilayer perceptron (MLP); it receives the statistical features or histogram of the node feature matrix X; the output is used to calculate the JS divergence or Wasserstein distance between the data and the real data distribution. The dynamic behavior branch analyzes node traffic sequences using a temporal neural network; the temporal neural network extracts the time and frequency domain features of the traffic sequences by integrating Fast Fourier Transform (FFT), and compares the differences between generated traffic and real traffic in terms of periodicity, fluctuation patterns, and suddenness. The output feature vectors of the topology branch, static attribute branch, and dynamic behavior branch are concatenated and then aggregated through a fully connected layer to generate the final authenticity score.

[0042] Furthermore, the loss function is established, including: The generator's total loss function L_G is a weighted sum of adversarial loss, node degree distribution loss, and device distribution loss; The generator's total loss function L_G is given by the following formula: L_G=E[log(1 D(G'))]+λ1 ||Degree(A') - Degree_real||2+λ2 JS(P_model ||P_real); Among them, the adversarial loss E[log(1 - D(G'))] is used to induce the generation of the topology deception discriminator; Node degree distribution loss λ1 ||Degree(A') - Degree_real||2, used to force the generated topology to have a node degree distribution similar to the real network; Equipment distributed loss λ2 JS(P_model || P_real) represents minimizing the JS divergence between the generated topology and the real topology in terms of device type distribution; The discriminator's loss function is the standard adversarial loss L_D, as shown in the following formula: L_D = -[E[log(D(G))] + E[log(1 - D(G'))]]; Where E[log(D(G))] is the expected probability of the real data being judged as true; E[log(1 - D(G'))] is the expected probability of the generated data being judged as false.

[0043] Furthermore, the generated topology data is converted into a network management model, including: According to the mapping rules, the topology diagram data is converted into management objects defined by the telecommunications management network or the YANG model; the mapping rules include node mapping and link mapping. Generate network management configuration files in XML or JSON format that can be recognized by the network management system.

[0044] Furthermore, node mapping includes mapping features in the feature matrix of topology graph structure data nodes to device attributes; Link mapping includes converting the connection relationships in the adjacency matrix of the topology graph structure into connection lines or cross connections between ports, and organizing them into a containment tree structure that describes parent-child relationships.

[0045] Furthermore, it also includes: Establish a session with the network management interface of the target simulation environment via the Netconf or TL1 protocol, specify the target configuration database using the network management configuration file, and initiate a transactional configuration push request. After receiving the configuration, the target simulation environment automatically creates the corresponding virtual devices and establishes connections, constructing a simulation network consistent with the topology data, which is used for operation and maintenance drills, algorithm testing, or capacity planning.

[0046] Thirdly, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, which stores computer programs; When a processor executes a computer program stored in memory, it implements the above-described method for generating a simulated topology for a telecommunications transmission network.

[0047] Fourthly, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the above-described method for generating a simulated topology of a telecommunications transmission network.

[0048] This disclosure includes at least the following beneficial effects: This disclosure deeply integrates telecommunications domain knowledge rules with generative GANs, resolving the dilemma of balancing realism and compliance in simulation topology generation. It can simultaneously generate highly realistic network structures, static configurations, and dynamic traffic. Through hard rule constraints, it ensures that the generated results conform to telecommunications network design specifications, such as dual core nodes and dual access aggregation. It achieves end-to-end one-click generation from data to a directly deployable simulation environment, significantly improving efficiency. The output is a standard network management model, seamlessly integrated with existing network management systems via the standard Netconf protocol, and is ready to use out of the box.

[0049] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the method flow for generating embodiments of this disclosure; Figure 2 This is a schematic diagram illustrating the principle of the generation method in the embodiments of this disclosure; Figure 3 This is a schematic diagram of the device structure for generating an embodiment of the present disclosure; Figure 4 This is a schematic diagram of the electronic device structure according to an embodiment of the present disclosure. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0053] like Figure 1 As shown, a method for generating a simulation topology for a telecommunications transmission network includes: S101: Obtain the engineering topology dataset of the telecommunications transmission network, extract multidimensional telecommunications features, construct conditional vectors, and generate standardized spatiotemporal graph data representations; S102, based on standardized spatiotemporal graph data representation, constructs a generator for generative adversarial networks and uses a multi-scale evaluation branch to construct a discriminator for generative adversarial networks; S103, Establish the loss function and train the generative adversarial network to obtain the trained network; S104, Generate topology graph structure data from the trained network based on the engineering topology dataset of the real-time telecommunications transmission network; S105 converts the generated topology diagram data into a network management model to obtain the network management configuration file.

[0054] In practical implementation, based on the principle Figure 2 The introduction is as follows: S101: Telecommunications Topology Data Encoding: This step aims to address the problem of transforming complex, heterogeneous real-world telecommunications network data into normalized, multimodal graph-structured data suitable for deep learning models.

[0055] S101.1 Obtain the engineering topology dataset of a real telecommunications transmission network and extract multi-dimensional telecommunications-specific features: S101.1.1, Data Source and Preprocessing: The engineering topology dataset originates from the configuration backup library or performance acquisition library of the existing network management system (NMS) and element management system (EMS). After data acquisition, data cleaning and alignment are required, including handling missing values, standardizing device naming conventions, and aligning data timestamps from different systems to form a consistent global network view.

[0056] S101.1.2, Specific details of feature extraction: In addition to device type, geographical level, and network level, network element node attributes can further include device model, rack slot capacity, number and speed of used ports, and device weight. These attributes collectively define the static functionality and capacity of the node.

[0057] Fiber connections are extended to link attributes, including physical fiber type, link length, logical link bandwidth, link cost, protection group identifier, and optical power attenuation value. These attributes define the physical and logical constraints of the connection.

[0058] Dynamic parameters include: service traffic matrix timing and protection switching event chain. The service traffic matrix timing is time-series data collected at fixed time granularities (e.g., 5 minutes) on network element-level inbound and outbound traffic, critical link utilization, and bandwidth timing of end-to-end service flows (e.g., LSP / tunnels), revealing the network's tidal effects and service patterns. The protection switching event chain is a structured log recording historical fault events, including event ID, timestamp, triggering network element / port, fault type (e.g., LOS, B1_EXC), associated protection group, switching status (success / failure), and switching duration. This data is used to learn the network's resilience behavior.

[0059] S101.2, Construct the dual-label conditional vector C: This step aims to transform discrete, symbolic network layer labels into a continuous, dense numerical vector. This vector can serve as conditional information to precisely control the output of the generative artificial intelligence (GAN), ensuring that the generated topology possesses the specified layer attributes.

[0060] S101.2.1, Dual-Label One-Hot Encoding Mapping For each network element node in the dataset, extract its two discrete category labels: geographical level label (e.g., province, city, district / county) and network level label (e.g., core, aggregation, access).

[0061] Perform one-hot encoding on both tags respectively: Assuming that the geographic level has L_geo possible values, a specific label (such as a city) is converted into a binary vector of length L_geo, where the position corresponding to "city" is 1 and the other positions are 0.

[0062] Similarly, assuming that the network layer has L_net possible values, the network layer label is converted into a binary vector of length L_net.

[0063] This process converts the two classification labels of each node into two sparse, high-dimensional binary vector representations.

[0064] S101.2.2, Sparse Feature Vector Concatenation and Dimensionality Reduction Mapping: The geographic-level one-hot vector and the network-level one-hot vector obtained from the above steps are concatenated along the feature dimension to form a longer sparse binary vector with a total dimension of L_geo + L_net.

[0065] Subsequently, a learnable embedding layer maps this high-dimensional sparse vector to a low-dimensional dense continuous vector space. This embedding layer is implemented as follows: Multilayer perceptron mapping: The concatenated sparse vector is directly input into a shallow multilayer perceptron (MLP), which contains a linear layer and a nonlinear activation function (such as ReLU). Its output is the desired low-dimensional dense vector.

[0066] The output of this sub-step is a dense floating-point vector of fixed-dimensional D_embed (such as 64-dimensional or 128-dimensional), which is called the initial condition vector of the node.

[0067] S101.2.3, Semantic optimization and output of condition vectors: The initial condition vector obtained in step S101.2.2 is used as a trainable parameter and is jointly optimized through backpropagation and gradient descent throughout the training process of the generative adversarial network (GAN).

[0068] The goal of optimization is to make the cosine or Euclidean distance between the conditional vectors of semantically similar hierarchical combinations (e.g., (province, core) and (city, convergence) since they both represent higher-level nodes in the network) closer, while the vector distance between combinations with large semantic differences (e.g., (province, core) and (district, access)) is farther.

[0069] Through this optimization, the resulting conditional vector C is no longer merely a simple mapping of labels, but a distributed representation that captures the deep semantic relationships between the hierarchical structures of the telecommunications network. It provides the generator with powerful and meaningful control signals, guiding it to generate a topology that conforms to the hierarchical rules of the telecommunications network.

[0070] S101.3, Generate standardized spatiotemporal graph data representation: G = (A, X, T), where: A represents the adjacency matrix (Aij=1 indicates that network elements i and j are connected by a fiber); it is a binary or multi-weighted square matrix. Aij=1 indicates the existence of a connection; in some embodiments, Aij can also be assigned the cost or weight of the link to carry more network information.

[0071] X represents the node feature matrix (Xi = [Device Type, Number of Ports, Geographical Level, Network Level]); it is a two-dimensional matrix, where each row of Xi represents the feature vector of a network element. Xi not only includes the one-hot encoding of the device type, the number of ports, etc., but also incorporates the condition vector C generated by step S101.2, so that the node features themselves carry hierarchical condition information.

[0072] T represents the time series matrix of dynamic parameters (Ti∈R) Let T represent the traffic sequence at time K for network element i. This is a flattened representation of a three-dimensional tensor. More precisely, for each node, its dynamic parameters Ti are a K × M matrix, where K is the time step and M is the number of dynamic features (such as inbound traffic, outbound traffic, CPU utilization, and memory utilization). The T matrix aggregates these time-series data from all nodes, collectively forming a dynamic behavioral profile of the network.

[0073] S102, Telecommunications Rule-Driven GAN Model Architecture Design: The core of this step lies in designing a generative adversarial network that incorporates prior knowledge from the telecommunications field, ensuring that the generated topology is not only statistically accurate but also structurally conforms to the design specifications of telecommunications networks.

[0074] S102.1, Generator G Design: Input: Noise vector Z + Condition vector C; The noise vector Z (sampled from a standard normal distribution) and the condition vector C are concatenated at the input layer.

[0075] Output: The simulated topology G' = (A', X', T') after rule correction and conformity to telecommunications standards; Telecommunications regulations, including: Full interconnection constraint for core layer nodes (enforcing A' core node degree ≥ threshold); traverse all nodes labeled "core" and check their degree. If the degree is less than the preset threshold (e.g., N-1, where N is the number of core nodes), activate a rule correction subroutine. This subroutine calculates the missing connections between current core nodes and adds connections with the highest priority to ensure the redundancy and reliability of the core network.

[0076] The access node is subject to a dual-homing constraint (forcing access node A' to have a degree of 2). It iterates through all nodes labeled "access" and checks their degree. If the degree is not equal to 2, another correction subroutine is activated: if the degree is less than 2, it preferentially connects to the same area's aggregation node; if the degree is greater than 2, it disconnects redundant connections, retaining only the two connections with the lowest cost. This simulates the topology of access rings or dual-homing uplinks in a real network.

[0077] S102.2, Discriminator Design (Multi-scale Evaluation Branch): Topological branches are used to represent the rationality of the verification hierarchy structure of the graph convolutional network; Static branches are used to statistically analyze the matching degree of network element type distribution (calculate the JS divergence of OTN device proportion). Dynamic branching is used to extract spectral features of the flow matrix (by comparing Fourier fundamental frequency energy).

[0078] The discriminator no longer judges only "true / false", but evaluates the realism of the generated topology from multiple dimensions. Its final decision is a combination of the evaluation results of multiple branches.

[0079] The topology branch is centered around a Graph Convolutional Network (GCN) or Graph Attention Network (GAT). It receives the topology G or G', aggregates neighbor information through a message-passing mechanism, and finally outputs a graph-level embedding vector. The purpose of this vector is to determine whether the connection pattern and hierarchical structure of the entire graph are consistent with the real telecommunications network.

[0080] The static attribute branch is a multilayer perceptron (MLP). It receives statistical features of the node feature matrix X (such as the quantity distribution of each device type, the mean / variance of the number of ports) or a histogram. Its output is used to calculate the JS divergence (Jensen-Shannon Divergence) or Wasserstein distance with the true data distribution to penalize static statistical differences such as device type imbalance.

[0081] The dynamic behavior branch uses temporal neural networks, such as one-dimensional convolutional neural networks (1D-CNN) or long short-term memory networks (LSTM), to analyze the node traffic sequence T. CNN / LSTM extracts the time-domain and frequency-domain features of the traffic sequence (by integrating a fast fourier transform (FFT) module) and compares the generated traffic with the real traffic in terms of periodicity, fluctuation patterns, and bursts.

[0082] Finally, the output feature vectors of the three branches are concatenated and aggregated through a fully connected layer to produce a final authenticity score.

[0083] S103, Multi-objective Joint Optimization Training Strategy: This step aims to simultaneously optimize the structure, static properties, and dynamic behavior of the topology through a carefully designed loss function and optimization techniques.

[0084] S103.1, Loss Function Design: L_G=E[log(1 D(G'))]+λ1 ||Degree(A') - Degree_real||2+λ2 JS(P_model ||P_real) The generator's total loss function L_G is a weighted sum of the following multiple losses: 1. Adversarial loss: E[log(1 - D(G'))], which prompts the generation of a topology deception discriminator.

[0085] 2. Node degree distribution loss: λ1 ||Degree(A') - Degree_real||2, forces the generated topology to have a node degree distribution (not just core and access nodes) similar to the real network.

[0086] 3. Equipment distribution loss: λ2 JS(P_model || P_real) minimizes the JS divergence between the generated topology and the real topology in terms of device type distribution.

[0087] The discriminator's loss function is the standard adversarial loss: L_D = -[E[log(D(G))] + E[log(1 - D(G'))]].

[0088] S103.2, Dynamic parameter timing optimization: Local patterns of the generated traffic sequence and the real sequence are aligned using DTW (Dynamic Time Warping). S104, seamless integration with the control system: This step solves the "last mile" problem in generating topology from data to a practically usable simulation environment, achieving closed-loop automation.

[0089] S104.1, Conversion of topology data to network management model: Define a mapping rule base to convert the generated graph structure data G' into managed objects defined by the Telecommunication Management Network (TMN) or the YANG model.

[0090] Node mapping: Maps features in X' to attributes of device MO (e.g., neId, neType, neLocation).

[0091] Link mapping: The connection relationships in A' are converted into connection lines MO (ctp) or cross-connect MO (xconnect) between ports, and organized into a Containment tree structure that describes the parent-child relationship. This structure is the basis for the network management system to understand the composition and connection of devices.

[0092] Finally, a complete configuration file in XML or JSON format that can be recognized by the network management system is generated.

[0093] S104.2, Lossless configuration injection for the simulation environment: Establish a session with the network management interface of the target simulation environment (such as GNS3, EVENG, or the vendor's own simulator) via the Netconf or TL1 protocol. Encapsulate the generated configuration file within the Netconf... <edit-config>During the operation, the target configuration database is specified, and a transactional configuration push request is initiated.

[0094] After receiving the configuration, the simulation environment automatically creates the corresponding virtual devices and establishes connections, thereby quickly building a simulation network that is completely consistent with the generated topology G' within minutes. This network can be directly monitored and managed by traditional network management systems for subsequent operation and maintenance drills, algorithm testing, or capacity planning.

[0095] This disclosure addresses issues related to topology generation, parameter system generation, control system adaptation, and simulation efficiency, specifically including: How can the generated simulation topology simultaneously satisfy the following requirements: reflecting regional differences (e.g., mountainous chain topology vs. urban mesh topology); and complying with telecommunications engineering rules (e.g., full interconnection of core nodes and dual aggregation of access nodes).

[0096] How to achieve this: Static parameters conform to the actual device distribution (such as the proportion of a certain type of router, the proportion of port types); dynamic parameters restore the timing characteristics of services (such as traffic periodic peaks, protection switching event chains).

[0097] How to seamlessly import the generated topology into the management system development environment (such as supporting Netconf / YANG model parsing) to avoid data loss caused by manual conversion.

[0098] How can we improve topology construction efficiency by more than 10 times by replacing manual repetitive design with the one-time training and multiple generation mechanism of GAN?

[0099] like Figure 3 As shown, a telecommunications transmission network simulation topology generation device includes: The system comprises a data construction unit 301, an adversarial network construction unit 302, a training unit 303, a topology generation unit 304, and a data conversion unit 305. Data construction unit 301 is used to acquire engineering topology datasets of telecommunications transmission networks, extract multidimensional telecommunications features, construct condition vectors, and generate standardized spatiotemporal graph data representations. The adversarial network building unit 302 is based on standardized spatiotemporal graph data representation and is used to build the generator of the generative adversarial network. It also uses a multi-scale evaluation branch to build the discriminator of the generative adversarial network. Training unit 303 is used to establish a loss function and train the generative adversarial network to obtain a trained network. Topology generation unit 304 is used to generate topology graph structure data from a trained network based on the engineering topology dataset of a real-time telecommunications transmission network. The data conversion unit 305 is used to convert the generated topology diagram structure data into a network management model to obtain a network management configuration file.

[0100] like Figure 4 As shown, this disclosure provides an electronic device, including a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. Memory 403 stores computer programs; The processor 401 implements the above method when executing a computer program stored in the memory 403.

[0101] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0102] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0103] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0104] To enable those skilled in the art to better understand this disclosure, the principles of this disclosure are explained below in conjunction with the accompanying drawings: Specific implementation of telecommunications topology data encoding: Data acquisition and feature extraction: The raw data contains 5,832 network elements and 12,451 link records.

[0105] Data cleaning: After removing test network elements, isolated connectionless network elements, and false alarm links, the effective dataset is obtained: N = 4215 network elements, L = 9687 links. The data validity rate is (4215+9687) / (5832+12451) ≈ 76.5%.

[0106] Feature statistics: Equipment type distribution: OTN: 18.2% (767 units), PTN: 81.8% (3448 units).

[0107] Network layer distribution: Core layer: 1.5% (63), Aggregation layer: 10.1% (426), Access layer: 88.4% (3726).

[0108] Average node degree: (2) 9687) / 4215 ≈ 4.6. (Calculation formula: <k>= 2L / N) Construct the condition vector C: The one-hot encodings of the geographic level (3 types) and the network level (3 types) are concatenated into a 6-dimensional sparse vector.

[0109] A (6, 16) embedding layer is used to map sparse vectors to 16-dimensional dense conditional vectors C_i. The embedding layer parameters are updated during training via backpropagation.

[0110] Construction of graph representation: Adjacency matrix A: 4215 x 4215 dimensions, density 9687 / (4215) (4215-1) / 2) ≈0.0011, which is a very sparse matrix.

[0111] Node feature matrix X: The feature vector of each node is [device type (2D), number of ports (1D), condition vector (16D)], so X is a 4215 x 19 matrix.

[0112] Time series matrix T: The original traffic data is 90 days, 5-minute granularity (25920 points). It is then downsampled to an hourly granularity (90... (24 = 2160 points), and apply Z-score standardization (to make the mean 0 and the standard deviation 1). The final dimension of T is 4215 x 2160.

[0113] GAN modeling and training (including hyperparameters and computation): Hyperparameter settings: Noise dimension dim_z = 100; The condition vector dimension dim_c = 16; The generator / discriminator learning rate lr = 1e-4; Batch size = 32; Optimizer: Adam (beta1=0.5, beta2=0.999); Loss weights: λ1 = 1.0 (degree distribution), λ2 = 5.0 (JS divergence), λ3 = 2.0 (DTW).

[0114] Rule constraint calculation: Core layer constraints: 63 core nodes; theoretically, the minimum number of links required for full interconnection is 63. (63-1) / 2 = 1953. The number of links between core nodes in the generator's original output A'_raw is 1721. After calculation and supplementation by the rule module, 1953 - 1721 = 232 missing links are added.

[0115] Access layer constraints: An inspection of 3726 access nodes revealed that 812 nodes had a degree other than 2. The rules module executed 1254 "delete redundant connections" operations and 688 "add missing connections" operations.

[0116] Discriminator multi-scale evaluation calculation: Static branching (JS divergence calculation): In the actual data, the proportion of OTN is P_real = 767 / 4215 ≈ 0.182.

[0117] In the generated batch of data, OTN accounted for 0.169% of the total.

[0118] JS divergence calculation: M = 0.5 (P_real + P_model) = 0.1755. JS(P_real || P_model) = 0.5 KL(P_real || M) + 0.5 KL(P_model || M) ≈ 0.00041. (KL is the KL divergence). Dynamic branching (DTW distance calculation - single-node example): A random access node is selected, and its generated traffic sequence T'_i and actual sequence T_i are both vectors of length 2160.

[0119] The FastDTW algorithm was used to calculate the alignment path and cumulative distance between the two sequences, resulting in DTW(T'_i, T_i) = 318.7.

[0120] In comparison, the calculated MSE loss for this node is 85.2. DTW is better at capturing phase differences.

[0121] Training results and performance quantification: The model was trained for 500 epochs, and the convergence of key metrics was recorded.

[0122] Loss function convergence: The generator loss L_G converges stably from the initial -12.5 to around -1.8.

[0123] The discriminator loss L_D stabilized around 0.1 from 0.5, indicating that the discriminator is still slightly better than the generator, but has reached a relative balance.

[0124] Generate quality assessment (round 500): Similarity of node degree distribution: The Jensen-Shannon divergence between the generated topology and the real topology is calculated to be 0.032 (the closer the value is to 0, the better; less than 0.05 is considered highly similar).

[0125] Device distribution matching degree: The JS divergence of the OTN / PTN device ratio is 0.008, indicating that the device composition of the generated topology is almost consistent with the real network.

[0126] Traffic similarity: 100 nodes were randomly selected, and the average DTW distance between their generated traffic and real traffic was calculated to be 285.4 ± 45.2. In contrast, the control group model trained using pure MSE loss had a DTW distance of 498.7 ± 112.6. The method disclosed in this paper improves the temporal similarity by 75%.

[0127] Integration performance assessment: Conversion to power: 100 generated topologies G' were converted into YANG models, all successfully, with no syntax or semantic errors.

[0128] Configure push performance: A complete topology configuration containing 4215 network elements is pushed to the simulation environment via the Netconf protocol.

[0129] Total time: 12 minutes and 35 seconds.

[0130] Breakdown of time consumption: Model inference generation 4.2 seconds, model conversion 1.8 seconds, configuration push 12 minutes and 29 seconds (network I / O is the main bottleneck).

[0131] Compared with traditional methods: Traditional manual configuration: Based on experience, it takes an average of 5 engineers to configure a topology of this size. (calculated at 8 hours per person per day, i.e., 2400 minutes).

[0132] Efficiency improvement: (2400 - 12.5) / 2400 100% ≈ 99.5%, an efficiency improvement of two orders of magnitude.

[0133] Resource costs: The main cost of this disclosure is the electricity cost for a one-time training session, while the labor cost far exceeds this.

[0134] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.< / k>

Claims

1. A method for generating a simulated topology for a telecommunications transmission network, characterized in that, The method includes: Obtain engineering topology datasets of telecommunications transmission networks, extract multidimensional telecommunications features, construct conditional vectors, and generate standardized spatiotemporal graph data representations; Based on standardized spatiotemporal graph data representation, a generator for generative adversarial networks (GANs) is constructed, and a discriminator for the GANs is constructed using a multi-scale evaluation branch. Establish a loss function and train the generative adversarial network to obtain the trained network; Based on the engineering topology dataset of real-time telecommunications transmission networks, topology graph structure data is generated through trained networks; The generated topology data is converted into a network management model to obtain the network management configuration file.

2. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, The engineering topology dataset for telecommunications transmission networks includes: Configuration backup library or performance data collection library for the existing network management system and network element management system.

3. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, Obtain the engineering topology dataset of the telecommunications transmission network and extract multidimensional telecommunications features, including: After data acquisition, data cleaning and alignment are performed, including handling missing values, standardizing device naming conventions, and aligning data timestamps from different system sources to form a consistent global network view.

4. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, Extracting multidimensional telecommunications Features include: network element node attributes, link attributes, service traffic matrix time-series data, and historical fault event data; Network element node attributes include: device type, geographical level, network level, device model, rack slot capacity, number of used ports and port speed, and device weight; Link attributes include: physical fiber type, link length, logical link bandwidth, link cost, protection group identifier, and optical power attenuation value; The service traffic matrix time series data includes: network element-level inbound and outbound traffic, critical link utilization, and bandwidth time series data of end-to-end service flows collected at fixed time granularity; Historical fault event data includes: event ID, timestamp, triggering network element and port, fault type, associated protection group, switching status, and switching duration.

5. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, Constructing the condition vector includes: For each network element node in the dataset, the category label is extracted, and one-hot encoding is performed on the category label. The category label of each network element node is converted into a high-dimensional one-hot vector. The category label includes: geographic level label and network level label. The high-dimensional one-hot vectors of different category labels are concatenated along the feature dimension to form a sparse vector; the sparse vector is then mapped to a low-dimensional dense continuous vector space through an embedding layer, and the initial condition vector of the node is output. The initial condition vector is used as a trainable parameter. During the entire training process of the generative adversarial network, the final condition vector is obtained through joint optimization using the backpropagation algorithm and gradient descent.

6. The method for generating a simulated topology for a telecommunications transmission network according to claim 5, characterized in that, The embedding layer uses a multilayer perceptron mapping, inputting sparse vectors into the multilayer perceptron MLP and outputting low-dimensional dense vectors as initial condition vectors; wherein, the MLP includes: a linear layer and a nonlinear activation function.

7. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, Generate standardized spatiotemporal graph data representations, including: Generate a standardized spatiotemporal graph data representation: G = (A, X, T), where: A represents the adjacency matrix; X represents the node feature matrix; T represents the dynamic parameter time sequence matrix.

8. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, Generators for building generative adversarial networks include: The generator's input is a noise vector Z plus a condition vector C; where the noise vector Z and the condition vector C are concatenated at the input layer. The generator outputs topology graph structure data G'= (A', X', T') constrained by telecommunications rules; Among them, telecommunications rule constraints include: full interconnection constraints of core layer nodes and dual aggregation constraints of access nodes; A' represents the adjacency matrix; X' represents the node feature matrix; T' represents the dynamic parameter time series matrix; The constraint for full interconnection of core layer nodes is that A' core node degree ≥ threshold; the constraint for dual aggregation of access nodes is that A' access node degree = set value.

9. The method for generating a simulated topology for a telecommunications transmission network according to claim 1, characterized in that, A discriminator for a generative adversarial network is constructed using a multi-scale evaluation branch, including: The discriminator includes: topology branch, static attribute branch, and dynamic behavior branch; The topology branch is based on the graph convolutional network GCN or the graph attention network GAT. It receives standardized spatiotemporal graph data representation or topology graph structure data, aggregates neighbor information through message passing mechanism, and finally outputs graph-level embedding vectors. The embedding vectors are used to determine whether the connection mode and hierarchical structure of the entire graph are consistent with the real telecommunications network. The static attribute branch is a multilayer perceptron (MLP); it receives the statistical features or histogram of the node feature matrix X; the output is used to calculate the JS divergence or Wasserstein distance between the data and the real data distribution. The dynamic behavior branch analyzes node traffic sequences using a temporal neural network; the temporal neural network extracts the time and frequency domain features of the traffic sequences by integrating Fast Fourier Transform (FFT), and compares the differences between generated traffic and real traffic in terms of periodicity, fluctuation patterns, and suddenness. The output feature vectors of the topology branch, static attribute branch, and dynamic behavior branch are concatenated and then aggregated through a fully connected layer to generate the final authenticity score.

10. The method for generating a simulated topology of a telecommunications transmission network according to claim 1, characterized in that, Establish the loss function, including: The generator's total loss function L_G is a weighted sum of adversarial loss, node degree distribution loss, and device distribution loss; The generator's total loss function L_G is given by the following formula: L_G=E[log(1 D(G'))]+λ1 ||Degree(A') - Degree_real||2+λ2 JS(P_model || P_real); Where D(G) represents the discriminator's evaluation output of the real topology; D(G') represents the discriminator's evaluation output of the generated topology; λ1, λ2, λ3 represent the weighted hyperparameters of the loss function; Degree(A') represents the sequence of node degrees of the generated topology; Degree_real represents the sequence of node degrees of the real topology; ||... ||2 represents the Euclidean norm; JS represents the Jensen-Shannon divergence; P_model represents the empirical distribution of device types in the generated topology; P_real represents the empirical distribution of device types in the real topology. The adversarial loss E[log(1 - D(G'))] is used to induce the generation of a topology deception discriminator; Node degree distribution loss λ1 ||Degree(A') - Degree_real||2, used to force the node degree distribution of the generated topology to be similar to that of the real network; Equipment distributed loss λ2 JS(P_model || P_real) represents minimizing the JS divergence between the generated topology and the real topology in terms of device type distribution; The discriminator's loss function is the standard adversarial loss L_D, as shown in the following formula: L_D = -[E[log(D(G))] + E[log(1 - D(G'))]]; Where E[log(D(G))] is the expected probability of the real data being judged as true; E[log(1 - D(G'))] is the expected probability of the generated data being judged as false.

11. The method for generating a simulated topology of a telecommunications transmission network according to claim 1, characterized in that, The generated topology data is converted into a network management model, including: According to the mapping rules, the topology diagram data is converted into management objects defined by the telecommunications management network or the YANG model; the mapping rules include node mapping and link mapping. Generate network management configuration files in XML or JSON format that can be recognized by the network management system.

12. The method for generating a simulated topology of a telecommunications transmission network according to claim 11, characterized in that, Node mapping includes: mapping features in the feature matrix of topology graph structure data nodes to device attributes; Link mapping includes converting the connection relationships in the adjacency matrix of the topology graph structure into connection lines or cross connections between ports, and organizing them into a containment tree structure that describes parent-child relationships.

13. The method for generating a simulated topology of a telecommunications transmission network according to claim 1, characterized in that, Also includes: Establish a session with the network management interface of the target simulation environment via the Netconf or TL1 protocol, specify the target configuration database using the network management configuration file, and initiate a transactional configuration push request. After receiving the configuration, the target simulation environment automatically creates the corresponding virtual devices and establishes connections, constructing a simulation network consistent with the topology data, which is used for operation and maintenance drills, algorithm testing, or capacity planning.

14. A telecommunications transmission network simulation topology generation device, characterized in that, include: Data construction unit, adversarial network construction unit, training unit, topology generation unit, and data transformation unit; The data construction unit is used to acquire engineering topology datasets of telecommunications transmission networks, extract multidimensional telecommunications features, construct conditional vectors, and generate standardized spatiotemporal graph data representations. The adversarial network building unit, based on standardized spatiotemporal graph data representation, is used to build the generator of the generative adversarial network, and a multi-scale evaluation branch is used to build the discriminator of the generative adversarial network. The training unit is used to establish the loss function and train the generative adversarial network to obtain the trained network. The topology generation unit is used to generate topology graph structure data from a trained network based on a real-time telecommunications transmission network engineering topology dataset. The data conversion unit is used to convert the generated topology diagram data into a network management model to obtain a network management configuration file.

15. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, The engineering topology dataset for telecommunications transmission networks includes: Configuration backup library or performance data collection library for the existing network management system and network element management system.

16. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Obtain the engineering topology dataset of the telecommunications transmission network and extract multidimensional telecommunications features, including: After data acquisition, data cleaning and alignment are performed, including handling missing values, standardizing device naming conventions, and aligning data timestamps from different system sources to form a consistent global network view.

17. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Extracting multidimensional telecommunications Features include: network element node attributes, link attributes, service traffic matrix time-series data, and historical fault event data; Network element node attributes include: device type, geographical level, network level, device model, rack slot capacity, number of used ports and port speed, and device weight; Link attributes include: physical fiber type, link length, logical link bandwidth, link cost, protection group identifier, and optical power attenuation value; The service traffic matrix time series data includes: network element-level inbound and outbound traffic, critical link utilization, and bandwidth time series data of end-to-end service flows collected at fixed time granularity; Historical fault event data includes: event ID, timestamp, triggering network element and port, fault type, associated protection group, switching status, and switching duration.

18. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Constructing the condition vector includes: For each network element node in the dataset, the category label is extracted, and one-hot encoding is performed on the category label. The category label of each network element node is converted into a high-dimensional one-hot vector. The category label includes: geographic level label and network level label. The high-dimensional one-hot vectors of different category labels are concatenated along the feature dimension to form a sparse vector; the sparse vector is then mapped to a low-dimensional dense continuous vector space through an embedding layer, and the initial condition vector of the node is output. The initial condition vector is used as a trainable parameter. During the entire training process of the generative adversarial network, the final condition vector is obtained through joint optimization using the backpropagation algorithm and gradient descent.

19. A telecommunications transmission network simulation topology generation device according to claim 18, characterized in that, The embedding layer uses a multilayer perceptron mapping, inputting sparse vectors into the multilayer perceptron MLP and outputting low-dimensional dense vectors as initial condition vectors; wherein, the MLP includes: a linear layer and a nonlinear activation function.

20. A telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Generate standardized spatiotemporal graph data representations, including: Generate a standardized spatiotemporal graph data representation: G = (A, X, T), where: A represents the adjacency matrix; X represents the node feature matrix; T represents the dynamic parameter time sequence matrix.

21. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Generators for building generative adversarial networks include: The generator's input is a noise vector Z plus a condition vector C; where the noise vector Z and the condition vector C are concatenated at the input layer. The generator outputs topology graph structure data G'= (A', X', T') constrained by telecommunications rules; Among them, telecommunications rule constraints include: full interconnection constraints of core layer nodes and dual aggregation constraints of access nodes; A' represents the adjacency matrix; X' represents the node feature matrix; T' represents the dynamic parameter time series matrix; The constraint for full interconnection of core layer nodes is that A' core node degree ≥ threshold; the constraint for dual aggregation of access nodes is that A' access node degree = set value.

22. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, A discriminator for a generative adversarial network is constructed using a multi-scale evaluation branch, including: The discriminator includes: topology branch, static attribute branch, and dynamic behavior branch; The topology branch is based on the graph convolutional network GCN or the graph attention network GAT. It receives standardized spatiotemporal graph data representation or topology graph structure data, aggregates neighbor information through message passing mechanism, and finally outputs graph-level embedding vectors. The embedding vectors are used to determine whether the connection mode and hierarchical structure of the entire graph are consistent with the real telecommunications network. The static attribute branch is a multilayer perceptron (MLP); it receives the statistical features or histogram of the node feature matrix X; the output is used to calculate the JS divergence or Wasserstein distance between the data and the real data distribution. The dynamic behavior branch analyzes node traffic sequences using a temporal neural network; the temporal neural network extracts the time and frequency domain features of the traffic sequences by integrating Fast Fourier Transform (FFT), and compares the differences between generated traffic and real traffic in terms of periodicity, fluctuation patterns, and suddenness. The output feature vectors of the topology branch, static attribute branch, and dynamic behavior branch are concatenated and then aggregated through a fully connected layer to generate the final authenticity score.

23. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Establish the loss function, including: The generator's total loss function L_G is a weighted sum of adversarial loss, node degree distribution loss, and device distribution loss; The generator's total loss function L_G is given by the following formula: L_G=E[log(1 D(G'))]+λ1 ||Degree(A') - Degree_real||2+λ2 JS(P_model || P_real); Where D(G) represents the discriminator's evaluation output of the real topology; D(G') represents the discriminator's evaluation output of the generated topology; λ1, λ2, λ3 represent the weighted hyperparameters of the loss function; Degree(A') represents the sequence of node degrees of the generated topology; Degree_real represents the sequence of node degrees of the real topology; ||... ||2 represents the Euclidean norm; JS represents the Jensen-Shannon divergence; P_model represents the empirical distribution of device types in the generated topology; P_real represents the empirical distribution of device types in the real topology. The adversarial loss E[log(1 - D(G'))] is used to induce the generation of a topology deception discriminator; Node degree distribution loss λ1 ||Degree(A') - Degree_real||2, used to force the node degree distribution of the generated topology to be similar to that of the real network; Equipment distributed loss λ2 JS(P_model || P_real) represents minimizing the JS divergence between the generated topology and the real topology in terms of device type distribution; The discriminator's loss function is the standard adversarial loss L_D, as shown in the following formula: L_D = -[E[log(D(G))] + E[log(1 - D(G'))]]; Where E[log(D(G))] is the expected probability of the real data being judged as true; E[log(1 - D(G'))] is the expected probability of the generated data being judged as false.

24. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, The generated topology data is converted into a network management model, including: According to the mapping rules, the topology diagram data is converted into management objects defined by the telecommunications management network or the YANG model; the mapping rules include node mapping and link mapping. Generate network management configuration files in XML or JSON format that can be recognized by the network management system.

25. A telecommunications transmission network simulation topology generation device according to claim 24, characterized in that, Node mapping includes: mapping features in the feature matrix of topology graph structure data nodes to device attributes; Link mapping includes converting the connection relationships in the adjacency matrix of the topology graph structure into connection lines or cross connections between ports, and organizing them into a containment tree structure that describes parent-child relationships.

26. The telecommunications transmission network simulation topology generation device according to claim 14, characterized in that, Also includes: Establish a session with the network management interface of the target simulation environment via the Netconf or TL1 protocol, specify the target configuration database using the network management configuration file, and initiate a transactional configuration push request. After receiving the configuration, the target simulation environment automatically creates the corresponding virtual devices and establishes connections, constructing a simulation network consistent with the topology data, which is used for operation and maintenance drills, algorithm testing, or capacity planning.

27. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, which stores computer programs; A processor, when executing a computer program stored in memory, implements a method for generating a simulated topology of a telecommunications transmission network according to any one of claims 1-13.

28. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for generating a simulation topology of a telecommunications transmission network as described in any one of claims 1-13.

Images