A method for OTN network routing optimization based on deep learning

Abstract

With the rapid development of information technology, OTN networks, as the backbone of modern communication networks, bear the burden of high-capacity, high-speed service transmission tasks. The purpose of researching OTN network routing optimization methods is to improve network performance and meet the ever-increasing communication demands. This invention introduces a deep learning-based OTN network routing optimization method. By selecting a deep learning algorithm model suitable for the characteristics of OTN networks, key network features are extracted, and intelligent routing decisions are achieved through model training and optimization. Results show that this method performs excellently in reducing transmission latency, improving bandwidth utilization, and reducing network congestion rates, significantly improving the performance of OTN networks and providing a new technical means for efficient network operation.

Description

Technical Field

[0001] This invention relates to the field of OTN (Optical Transport Network) technology, specifically to a method for intelligent routing calculation and path optimization of optical transport networks (OTN) using deep learning technology. Background Technology

[0002] Optical Transport Network (OTN) is a core technology of modern communication backbone networks, carrying massive data transmission tasks. In OTN networks, routing directly determines the transmission latency of services, the utilization rate of network resources, and the survivability of the network.

[0003] Existing OTN routing optimization technologies primarily rely on traditional graph theory algorithms, such as Dijkstra's algorithm, K-shortest path algorithm (KSP), or Constrained Shortest Path First (CSPF) algorithm. However, with the expansion of network scale and the diversification of service types (such as 5G slicing and computing power networks), existing technologies face the following challenges:

[0004] 1. High computational complexity: In large-scale mesh network topologies, as constraints (such as latency, dispersion, and OSNR) increase, the computational complexity of traditional algorithms increases exponentially, making it difficult to meet the requirements for real-time service activation.

[0005] 2. Lack of predictive ability: Traditional algorithms only calculate based on the current network status (such as the remaining bandwidth of the link), and cannot perceive future traffic trends, which can easily lead to local link congestion and cause network instability.

[0006] 3. Insufficient consideration of nonlinear impairments: Traditional routing algorithms usually simplify physical impairments (such as nonlinear effects) to link weights, which makes it difficult to accurately reflect the true transmission quality of optical signals under complex paths.

[0007] Therefore, a deep learning method is needed that can utilize historical data features, has fast reasoning capabilities, and can comprehensively consider physical damage to optimize OTN network routing. Summary of the Invention

[0008] The purpose of this invention is to provide a method for OTN network routing optimization based on deep learning. By constructing a deep neural network model, the nonlinear mapping relationship between network topology and service requirements is learned, thereby achieving millisecond-level intelligent routing decisions.

[0009] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for OTN network routing optimization based on deep learning, comprising the following steps:

[0010] Step 1: Collect and preprocess raw data on OTN network topology, links, and traffic to construct a standardized model input tensor;

[0011] Step 2: Based on the standardized graph tensor and temporal flow data output in Step 1, spatial topological features are extracted using a graph neural network and temporal flow features are extracted using a recurrent neural network, respectively, to complete the spatiotemporal feature fusion.

[0012] Step 3: Based on the spatiotemporal fusion feature vector output in Step 2, input the pre-trained routing decision model and output the link recommendation weight matrix or the optimal routing path;

[0013] Step 4: Based on the link recommendation weights output in Step 3, and combined with traditional graph theory algorithms, generate the final executable route and send it to the optical cross-connect device to complete the OTN service configuration;

[0014] Step 5: Based on the operational feedback data after the business configuration in Step 4, perform online incremental learning and iteratively update the routing decision model parameters.

[0015] In some possible implementations, step 1 specifically involves:

[0016] Step 1.1, Raw Data Acquisition: The SDN controller periodically collects network-wide optical performance monitoring (OPM) data to obtain two types of core raw data:

[0017] ① Topology data: Abstracting the OTN network as an undirected graph Collect raw information about nodes and links; node characteristics include node ID, node type, number of remaining ports, and current processing latency; link characteristics include link length, number of remaining spectrum slots, signal-to-noise ratio (OSNR) margin, and historical average load.

[0018] ② Traffic data: Collect historical data on the distribution of source and destination nodes for service requests, bandwidth requirements, and time-series data on service arrival intervals;

[0019] Step 1.2, Data Standardization Preprocessing: Normalize the features of all nodes and links to generate a feature matrix. Adjacency matrix with added self-loops Degree matrix We construct a standardized tensor that can be input into a deep learning model, which serves as the basic input for subsequent spatiotemporal feature extraction.

[0020] Step 1: Abstract the OTN as an undirected graph. By collecting multi-source data such as node and link optical performance and historical traffic through SDN, a feature matrix is ​​generated after normalization. Self-loop adjacency matrix Degree matrix The input is converted into a tensor that can be used in the model. This method fully preserves optical layer constraints such as OSNR and spectrum, avoiding the shortcomings of traditional routing that only focuses on bandwidth. Normalization eliminates dimensional differences, accelerates model convergence, improves prediction accuracy, and standardizes graph tensors to adapt to GCN input, reducing the overhead of large-scale topology calculations.

[0021] In some possible implementations, step 2 specifically involves:

[0022] Step 2.1, Spatial Topological Feature Extraction: Using the output from Step 1 , , As input, a multi-layered GCN graph convolutional layer is used to aggregate the association information of all neighbor nodes in the network. The propagation rule formula for the graph convolutional layer is:

[0023]

[0024] in, For trainable weight matrix, The ReLU activation function is used; the network topology high-dimensional space embedding vector is output layer by layer iteratively to capture the spatial correlation of nodes and links;

[0025] Step 2.2, Traffic Temporal Feature Extraction: Based on the historical traffic temporal tensor output in Step 1, LSTM / TCN is used to extract the spatiotemporal distribution statistical features of traffic; the spatial embedding vector output by GCN is connected to the gated recurrent unit (GRU) to capture the dynamic changes of traffic within a continuous time slice, predict the link load trend at the next moment, and output the temporal feature vector.

[0026] Step 2.3, Spatiotemporal Feature Fusion: The output spatial embedding vector from Step 2.1 and the output traffic temporal feature vector from Step 2.2 are concatenated and fused to generate a unified spatiotemporal fusion feature vector.

[0027] Step 2 takes the tensor from Step 1 as input, uses the graph convolution propagation rule formula to aggregate topological spatial relationships through GCN, and combines LSTM / GRU to extract traffic temporal features and predict link load. The two types of features are fused to obtain a unified spatiotemporal vector. It breaks through the limitations of Dijkstra and KSP static weighting, globally quantizes the influence of optical nonlinear damage, avoids congestion oscillations caused by instantaneous routing by relying on traffic prediction, and reduces the amount of computation by lightweight graph feature extraction, supporting millisecond-level online inference.

[0028] In some possible implementations, step 3 specifically involves:

[0029] Step 3.1, Building the routing decision model architecture: A hybrid architecture of GraphSAGE / Transformer / ST-GCN with attention mechanism is adopted. The input is the spatiotemporal fusion feature vector from Step 2, and the model outputs the link congestion probability, path QoT transmission quality score, and link recommendation weight.

[0030] Step 3.2, Offline Model Training: Using historical network-wide running data as the training set, a weighted composite loss function is constructed to complete the iterative update of parameters. The loss function formula is:

[0031]

[0032] in, The penalty for business disruption For link load variance, Penalties for substandard transmission quality These are hyperparameters; the training objective is to minimize the service blocking rate, balance the load on all network links, and meet the OSNR (Optical Signal-to-Noise Ratio) constraint. All weight parameters of the model are updated through backpropagation to obtain the pre-trained routing decision model.

[0033] Step 3.3, Online Real-time Inference: When a new service request arrives, the current network snapshot tensor collected and preprocessed in Step 1 is sent to the pre-trained model to quickly infer and output the recommended weight matrix of the entire network link or the candidate optimal routing path.

[0034] Step 3 feeds the fused features into the ST-GCN model and uses a composite loss function to complete offline training with the goals of low blocking, load balancing, and meeting QoT. When a new service arrives, forward inference outputs the link weights or the optimal path. The multi-objective loss function is adapted to the multi-constraint requirements of 5G and computing networks, reducing the traditional exponential computation to linear and significantly improving the inference speed. The dynamic link weight balances the entire network spectrum and incorporates optical impairments into the penalty term, solving the problem of path quality assessment distortion in traditional algorithms.

[0035] In some possible implementations, step 4 specifically involves:

[0036] Step 4.1, Route path generation: The route calculation module loads the link recommendation weights output in step 3 and runs the Dijkstra algorithm to solve for the optimal transmission path;

[0037] Step 4.2, Device Configuration Distribution: The optimal routing path is distributed to the OXC / OTN Switch optical cross-connect device to complete the cross-connect configuration and service activation of the OTN service optical channel.

[0038] Step 4 replaces the fixed cost with the dynamic link weight output by the model, runs Dijkstra to generate optical channels, and distributes them to OXC / OTN switches to complete the service configuration. The solution is compatible with existing SDN and optical transmission hardware, requires low investment in transformation, and the dynamic weight takes into account load, congestion risk and optical performance, effectively reducing service blockage, adapting to existing operation and maintenance processes, and shortening the service activation time.

[0039] In some possible implementations, step 5 specifically involves:

[0040] Step 5.1, Operational Status Feedback Collection: Collect the operational results after service activation in Step 4, including service establishment success / failure status, actual link load, real-time OSNR, transmission latency, and QoT actual transmission quality feedback data;

[0041] Step 5.2, Online Loss Calculation: Input the feedback of the real network state into the composite loss function and recalculate the real-time loss value. ;

[0042] Step 5.3, Incremental update of model parameters: Based on the real-time loss value, backpropagation is used to fine-tune the weight parameters of the routing decision model to achieve online incremental learning, adapt to long-term dynamic changes in network topology and traffic, and optimize the accuracy of subsequent routing inference.

[0043] Step 5 collects feedback data such as actual service load, OSNR, and blocking results, substitutes them into the loss function to calculate real-time loss, and fine-tunes model parameters to adapt to long-term changes in topology, traffic, and fiber status. Closed-loop self-optimization eliminates the need for full offline retraining, saving computing power and time, continuously correcting routing deviations, and optimizing various network indicators in a long-term stable manner, thereby reducing model operation and maintenance costs.

[0044] The beneficial effects of this invention are: This invention forms a complete closed loop of data preprocessing—spatiotemporal feature extraction—multi-objective deep learning inference—service configuration—online iterative optimization, which solves the shortcomings of traditional routing calculations, lack of traffic prediction, and insufficient consideration of optical impairment. It achieves network operation effects with low blocking, load balancing, high spectrum utilization, and low latency. The millisecond-level routing capability is adapted to new high-speed services, and it is compatible with existing equipment, has low operation and maintenance costs, and is easy to deploy on a large scale for commercial use. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the overall process architecture of the method of the present invention.

[0046] Figure 2 This is a framework diagram illustrating the interaction between a deep reinforcement learning agent and the OTN network environment. Detailed Implementation

[0047] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0048] like Figure 1 and 2 As shown, this invention provides a method for OTN network routing optimization based on deep learning, comprising the following steps:

[0049] 1. Network Topology Graph Embedding: The physical topology (nodes and links) of the OTN network is transformed into a graph data structure. Graph Neural Network (GNN) is used to extract high-dimensional feature vectors of nodes and links to generate a network state embedding representation.

[0050] 2. Business traffic feature extraction: Collect the source and destination node distribution, bandwidth requirements and arrival time intervals of historical business requests, and use Long Short-Term Memory Network (LSTM) or Temporal Convolutional Network (TCN) to extract the spatiotemporal distribution features of traffic.

[0051] 3. Construct a route prediction model: Design an end-to-end deep neural network model (such as GraphSAGE or Transformer architecture based on attention mechanism) that takes network state embedding and traffic features as input and outputs the congestion probability of the link or the QoT (quality of transport) score of the path.

[0052] 4. Model Training and Optimization: Using historical network operation data as the training set, the network parameters are updated using the backpropagation algorithm with the objective function of minimizing network congestion rate, maximizing spectrum utilization, or minimizing transmission latency.

[0053] 5. Online route reasoning: When a new service request arrives, the current network snapshot is input into the trained deep learning model, and the model quickly outputs the recommended optimal route path or link weight matrix to guide the configuration of the optical cross-connect device (OXC / OTNSwitch).

[0054] Working principle

[0055] Deep learning algorithm model selection

[0056] In OTN network routing optimization, selecting a suitable deep learning algorithm model is crucial for improving routing decision performance. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), as important branches of deep learning, have demonstrated significant advantages in routing optimization problems. CNNs, through local connections and shared weight mechanisms, can effectively process network data with grid-like topologies, such as network topology graphs and traffic distribution matrices. This characteristic makes CNNs excellent at extracting network spatial features, especially suitable for routing optimization tasks requiring global analysis of the network structure. On the other hand, Recurrent Neural Networks and their variant, Long Short-Term Memory (LSTM), excel at processing time-series data, capturing dynamically changing traffic patterns and link state information in the network. This is of great significance for the frequently changing service requirements and real-time routing adjustments in OTN networks. Combining the advantages of both, this invention adopts a hybrid model architecture, utilizing CNNs to extract spatial features of the network while simultaneously using LSTMs to handle dynamic changes in the time dimension, thereby achieving comprehensive perception and intelligent decision-making in complex network environments.

[0057] The selection of the aforementioned models is primarily based on their successful applications and theoretical advantages in relevant fields. Firstly, the superior performance of CNNs in image recognition and graphics data processing has been widely validated, and since network topology can be abstracted as graph data, CNNs can efficiently learn the implicit patterns within the network. Secondly, the unique design of LSTM in addressing temporal dependencies allows it to effectively handle the non-stationarity and burstiness of network traffic, avoiding gradient vanishing or exploding problems caused by long-term dependencies. Furthermore, research on routing algorithms based on deep reinforcement learning has shown that RNN series models have significant advantages in handling high-dimensional continuous state spaces, providing more accurate action value predictions for routing optimization. In summary, the model selected in this invention not only fully utilizes the spatial and temporal characteristics of the network but also possesses strong generalization ability and adaptability, laying a solid theoretical foundation for OTN network routing optimization.

[0058] Network Feature Extraction

[0059] Network feature extraction is a core component of deep learning-based OTN network routing optimization methods, aiming to extract crucial information for routing decisions from complex network environments. This invention identifies three key features: network topology, traffic distribution, and link status. Network topology reflects the connectivity between nodes and is one of the fundamental inputs for routing calculations. By encoding the topology as graph data and using Graph Convolutional Neural Networks (GCNs) for feature extraction, the spatial correlation between nodes and its impact on routing path selection can be effectively captured. Traffic distribution describes the load on different links in the network, typically presented in time series form. Statistical analysis of historical traffic data can extract statistical features such as peak, average, and fluctuation ranges, providing real-time load information for the routing algorithm. Link status includes parameters such as available bandwidth, latency, and packet loss rate, which directly affect the quality assessment of routing paths. By periodically monitoring and normalizing link status, it can be transformed into acceptable input features for the model.

[0060] The feature extraction method is designed with full consideration of the complexity and dynamism of OTN networks. For network topology, this invention employs a graph embedding-based technique to map high-dimensional topology information into a low-dimensional vector space, thereby reducing computational complexity while preserving key structural information. For traffic distribution, this invention combines a sliding window mechanism and statistical analysis methods to extract traffic features at multiple time scales to address routing requirements at different time granularities. For link status, this invention introduces a method combining online monitoring and offline analysis to ensure the real-time nature and accuracy of the feature data. The importance of these feature extraction methods lies in the fact that they not only provide high-quality input data for deep learning models but also lay a reliable foundation for subsequent route optimization. By comprehensively analyzing the above features, the model can gain a more comprehensive understanding of the network status, thereby generating more intelligent and efficient routing decisions.

[0061] Example

[0062] 1. Data Preprocessing and Input Construction

[0063] In this embodiment, the OTN network is modeled as an undirected graph. .

[0064] Node characteristics include node ID, node type (OXC / OADM), number of remaining ports, and current processing latency.

[0065] Link characteristics include link length, number of remaining spectrum slots, signal-to-noise ratio (OSNR) margin, and historical average load.

[0066] Input construction: After normalizing the above features, a feature matrix is ​​constructed. and adjacency matrix , which serves as the input to the graph convolutional network.

[0067] 2. Deep Learning Model Design

[0068] This invention proposes a spatiotemporal graph convolutional network (ST-GCN) architecture:

[0069] Spatial dependency capture: This involves aggregating neighbor node information using multi-layer graph convolutional layers (GCN layers). For the first... The propagation rules of a layer are defined as follows:

[0070] ;

[0071] in, To add a self-loop to the adjacency matrix, For trainable weight matrix, This is the ReLU activation function. This step is used to capture the spatial correlation of the network topology.

[0072] Time-dependent capture: A gated cyclic unit (GRU) is added after the GCN layer to process traffic changes within a continuous time slice and predict the load trend of each link in the next moment.

[0073] Output layer: The probability score of each link being selected is output through the fully connected layer, and the final route is generated by combining the K-shortest path algorithm.

[0074] 3. Loss Function Design

[0075] To balance resource utilization and business quality, a composite loss function is designed:

[0076]

[0077] in, The penalty for business disruption This is for link load variance (to encourage load balancing). Penalties for substandard transmission quality This is a hyperparameter.

[0078] 4. Implementation Process

[0079] Step 1: The SDN controller periodically collects optical performance monitoring (OPM) data for the entire network.

[0080] Step 2: The data preprocessing module converts the data into tensors and inputs them into the inference engine.

[0081] Step 3: The inference engine outputs the "recommendation weight" for each link.

[0082] Step 4: The route calculation module runs the Dijkstra algorithm based on the recommendation weights to generate paths.

[0083] Step 5: Distribute the configuration and collect feedback for online incremental learning of the model to adapt to dynamic changes in network topology.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing OTN network routing based on deep learning, characterized in that, The steps include the following: Step 1: Collect and preprocess raw data on OTN network topology, links, and traffic to construct a standardized model input tensor; Step 2: Based on the standardized graph tensor and temporal flow data output in Step 1, spatial topological features are extracted using a graph neural network and temporal flow features are extracted using a recurrent neural network, respectively, to complete the spatiotemporal feature fusion. Step 3: Based on the spatiotemporal fusion feature vector output in Step 2, input the pre-trained routing decision model and output the link recommendation weight matrix or the optimal routing path; Step 4: Based on the link recommendation weights output in Step 3, and combined with traditional graph theory algorithms, generate the final executable route and send it to the optical cross-connect device to complete the OTN service configuration; Step 5: Based on the operational feedback data after the business configuration in Step 4, perform online incremental learning and iteratively update the routing decision model parameters.

2. The method for OTN network routing optimization based on deep learning according to claim 1, characterized in that, Step 1 is as follows: Step 1.1, Raw Data Acquisition: The SDN controller periodically collects network-wide optical performance monitoring (OPM) data to obtain two types of core raw data: ① Topology data: Abstracting the OTN network as an undirected graph Collect raw information about nodes and links; node characteristics include node ID, node type, number of remaining ports, and current processing latency; link characteristics include link length, number of remaining spectrum slots, signal-to-noise ratio (OSNR) margin, and historical average load. ② Traffic data: Collect historical data on the distribution of source and destination nodes for service requests, bandwidth requirements, and time-series data on service arrival intervals; Step 1.2, Data Standardization Preprocessing: Normalize the features of all nodes and links to generate a feature matrix. Adjacency matrix with added self-loops Degree matrix We construct a standardized tensor that can be input into a deep learning model, which serves as the basic input for subsequent spatiotemporal feature extraction.

3. The method for OTN network routing optimization based on deep learning according to claim 2, characterized in that, Step 2 is as follows: Step 2.1, Spatial Topological Feature Extraction: Using the output from Step 1 , , As input, a multi-layered GCN graph convolutional layer is used to aggregate the association information of all neighbor nodes in the network. The propagation rule formula for the graph convolutional layer is: in, For trainable weight matrix, The ReLU activation function is used; the network topology high-dimensional space embedding vector is output layer by layer iteratively to capture the spatial correlation of nodes and links; Step 2.2, Traffic Temporal Feature Extraction: Based on the historical traffic temporal tensor output in Step 1, LSTM / TCN is used to extract the spatiotemporal distribution statistical features of traffic; the spatial embedding vector output by GCN is connected to the gated recurrent unit (GRU) to capture the dynamic changes of traffic within a continuous time slice, predict the link load trend at the next moment, and output the temporal feature vector. Step 2.3, Spatiotemporal Feature Fusion: The output spatial embedding vector from Step 2.1 and the output traffic temporal feature vector from Step 2.2 are concatenated and fused to generate a unified spatiotemporal fusion feature vector.

4. The method for OTN network routing optimization based on deep learning according to claim 3, characterized in that, Step 3 specifically involves: Step 3.1, Building the routing decision model architecture: A hybrid architecture of GraphSAGE / Transformer / ST-GCN with attention mechanism is adopted. The input is the spatiotemporal fusion feature vector from Step 2, and the model outputs the link congestion probability, path QoT transmission quality score, and link recommendation weight. Step 3.2, Offline Model Training: Using historical network-wide running data as the training set, a weighted composite loss function is constructed to complete the iterative update of parameters. The loss function formula is: in, The penalty for business disruption For link load variance, Penalties for substandard transmission quality These are hyperparameters; the training objective is to minimize the service blocking rate, balance the load on all network links, and meet the OSNR (Optical Signal-to-Noise Ratio) constraint. All weight parameters of the model are updated through backpropagation to obtain the pre-trained routing decision model. Step 3.3, Online Real-time Inference: When a new service request arrives, the current network snapshot tensor collected and preprocessed in Step 1 is sent to the pre-trained model to quickly infer and output the recommended weight matrix of the entire network link or the candidate optimal routing path.

5. The method for OTN network routing optimization based on deep learning according to claim 4, characterized in that, Step 4 specifically involves: Step 4.1, Route path generation: The route calculation module loads the link recommendation weights output in step 3 and runs the Dijkstra algorithm to solve for the optimal transmission path; Step 4.2, Device Configuration Distribution: The optimal routing path is distributed to the OXC / OTN Switch optical cross-connect device to complete the cross-connect configuration and service activation of the OTN service optical channel.

6. The method for OTN network routing optimization based on deep learning according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1, Operational Status Feedback Collection: Collect the operational results after service activation in Step 4, including service establishment success / failure status, actual link load, real-time OSNR, transmission latency, and QoT actual transmission quality feedback data; Step 5.2, Online Loss Calculation: Input the feedback of the real network state into the composite loss function and recalculate the real-time loss value. ; Step 5.3, Incremental update of model parameters: Based on the real-time loss value, backpropagation is used to fine-tune the weight parameters of the routing decision model to achieve online incremental learning, adapt to long-term dynamic changes in network topology and traffic, and optimize the accuracy of subsequent routing inference.

Images