Dynamic optimization method and device based on heterogeneous network topology, equipment and storage medium

Abstract

The invention discloses a dynamic optimization method, device and equipment based on heterogeneous network topology and a storage medium, which are applied to the technical field of network optimization, and the method comprises the following steps: obtaining equipment state data of a target network; processing the equipment state data to obtain equipment fingerprint data; processing the equipment state data and the equipment fingerprint data by using an incremental maximum likelihood topological reasoning technology to obtain a real-time topological graph; inputting the real-time topological graph into a fault diagnosis model for processing to obtain a root cause analysis result; performing state modeling processing on the traffic time sequence data, the equipment fingerprint data and the root cause analysis result by using a generative technology to obtain a deduction result; and performing strategy optimization processing on the deduction result by using a multi-agent reinforcement learning technology to obtain a network dynamic optimization scheme. The method provided by the invention can solve the technical problems in the prior art that the network management response is lagged, the flexibility is poor, and the global state is difficult to perceive in real time and the fault is difficult to locate accurately.

Description

A method, apparatus, device, and storage medium for dynamic optimization based on heterogeneous network topology. Technical Field

[0001] This invention relates to the field of network optimization technology, and in particular to a dynamic optimization method, apparatus, device, and storage medium based on heterogeneous network topology. Background Technology

[0002] With the rapid development of technologies such as the Internet of Things and 4G / 5G, the scale of heterogeneous networks (networks that integrate multiple communication technologies, device types and protocols) has expanded dramatically. Their topology changes frequently and their fault types are complex and diverse, such as link congestion, device downtime, and protocol conflicts.

[0003] In existing technologies, network management methods rely on manual configuration or static scanning, which are implemented through preset fixed rules and periodic polling monitoring. However, in the process of implementation, there are problems such as slow response, poor flexibility, and difficulty in perceiving the global real-time status. This leads to slow fault location, soaring operation and maintenance costs, degraded network performance, and difficulty in dealing with sudden traffic and new types of faults. Summary of the Invention

[0004] This invention provides a dynamic optimization method, apparatus, device, and storage medium based on heterogeneous network topology, which solves the technical problems of slow network management response, poor flexibility, difficulty in real-time perception of global status and accurate fault location in the prior art. It achieves the effects of real-time network topology inference, accurate fault root cause analysis, network status prediction and deduction, and dynamic autonomous optimization.

[0005] To address the aforementioned technical problems, this invention provides a dynamic optimization method based on heterogeneous network topology. The method includes: acquiring device status data of a target network; performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; processing the device status data and the device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology graph; and inputting the real-time topology graph into a fault diagnosis model constructed from a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network. The processing is configured to use the real-time topology graph as a static graph structure. Based on the device nodes and network links of the target network, a multimodal feature matrix is ​​constructed on the static graph structure. Spatiotemporal graph convolutional network technology is used to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network. Traffic time-series data of the target network is acquired, and generative techniques are used to perform state modeling processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network. Multi-agent reinforcement learning technology is used to optimize the inference results to obtain a dynamic optimization scheme for the network.

[0006] As one preferred embodiment, the step of performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data includes: performing feature vectorization on the device status data to obtain a high-dimensional feature vector; performing dimensionality reduction and privacy protection processing on the high-dimensional feature vector using a preset sparse random projection matrix to obtain a dimensionality-reduced binary signature; performing equal segmentation on the binary signature; and performing concatenation processing based on the obtained several signature segments to obtain the device fingerprint data.

[0007] As a preferred embodiment, the step of processing the device status data and device fingerprint data using incremental maximum likelihood topological reasoning technology to obtain a real-time topological graph includes: performing distributed information exchange processing on the device status data and device fingerprint data to obtain global neighbor relationship tuples and device fingerprint mapping data; processing the global neighbor relationship tuples using incremental maximum likelihood topological reasoning technology to obtain a topological graph structure; and obtaining the real-time topological graph based on the device fingerprint mapping data and the topological graph structure.

[0008] As one preferred embodiment, the step of using spatiotemporal graph convolutional network technology to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network includes: using a graph convolutional network to extract spatial features from the multimodal feature matrix to obtain spatial feature data; using a temporal convolutional network to extract temporal features from the multimodal feature matrix to obtain temporal feature data; fusing the spatial feature data and the temporal feature data to obtain fused high-order feature data; and performing root cause localization processing on the fused high-order feature data to obtain the root cause analysis results.

[0009] As one preferred embodiment, the step of using generative techniques to perform state modeling processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network includes: performing synchronous mapping processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain digital twin data; sequentially performing dynamic topology and multimodal state reconstruction processing on the digital twin data to obtain a network state snapshot; and sequentially performing state completion and simulation inference processing on the network state snapshot using generative techniques to obtain the inference results of the target network.

[0010] As one preferred embodiment, the step of using multi-agent reinforcement learning technology to optimize the inference results and obtain a network dynamic optimization scheme includes: using multi-agent reinforcement learning technology to optimize the inference results and obtain a joint policy model; and performing optimal policy generation processing on the joint policy model to obtain the network dynamic optimization scheme.

[0011] As one preferred embodiment, after obtaining the network dynamic optimization scheme, the dynamic optimization method based on heterogeneous network topology further includes: executing the network dynamic optimization scheme to obtain network status feedback data; performing difference comparison processing on the network status feedback data to obtain a simulation difference report; and performing cognitive model optimization processing on the simulation difference report to obtain the optimized network dynamic optimization scheme.

[0012] This invention also provides a dynamic optimization device based on heterogeneous network topology, comprising: an acquisition module for acquiring device status data of a target network; a first processing module for performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; processing the device status data and the device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology graph; and a second processing module for inputting the real-time topology graph into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network, wherein the processing is configured to use the real-time topology graph as a static graph structure. A multimodal feature matrix is ​​constructed on the static graph structure based on the device nodes and network links of the target network. Spatiotemporal graph convolutional network technology is used to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network. A modeling module is used to acquire the traffic time-series data of the target network and use generative techniques to perform state modeling processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network. An optimization module is used to use multi-agent reinforcement learning technology to optimize the inference results and obtain a dynamic optimization scheme for the network.

[0013] The present invention also provides a dynamic optimization device based on heterogeneous network topology, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the dynamic optimization method based on heterogeneous network topology as described above.

[0014] The present invention further provides a computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the dynamic optimization method based on heterogeneous network topology as described above.

[0015] Compared to existing technologies, the beneficial effects of this invention are at least one of the following: This invention acquires device status data of a target network; performs protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; processes the device status data and the device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology graph; and inputs the real-time topology graph into a fault diagnosis model constructed from a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network. The processing is configured to use the real-time topology graph as a static graph structure, based on... The device nodes and network links of the target network are used to construct a multimodal feature matrix attached to the static graph structure. Spatiotemporal graph convolutional network technology is used to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network. Traffic time-series data of the target network is acquired, and generative techniques are used to perform state modeling processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network. Multi-agent reinforcement learning technology is used to optimize the inference results to obtain a dynamic optimization scheme for the network.

[0016] Compared with existing technologies, this invention first extracts unified features from multi-source heterogeneous device status data and generates device fingerprint data through protocol-insensitive sparse feature hashing. Then, it uses incremental maximum likelihood topology inference technology to fuse device status data and device fingerprint data, replacing full static scanning with incremental updates to generate a real-time topology map. Subsequently, the real-time topology map is input into a fault diagnosis model constructed by a spatiotemporal graph convolutional network. Using this topology map as a static graph structure, an attached multimodal feature matrix is ​​constructed to extract global network correlation features from the spatial dimension and capture the temporal evolution law of the state from the temporal dimension. At the same time, combined with the traffic time series data of the target network, generative technology is used to fuse device fingerprint data and fault root cause analysis results to carry out network state modeling and inference. Finally, multi-agent reinforcement learning technology is used to optimize the state inference results. Based on the real-time global state of the network and future trend prediction, an optimization scheme adapted to the current network topology, device status, and traffic characteristics is dynamically generated. This adapts to the cross-node and cross-domain characteristics of heterogeneous networks, enhances the global dynamic adaptation capability and operational flexibility of network management, further shortens the fault handling response time, and achieves precise fault handling. Attached Figure Description

[0017] Figure 1 is a flowchart illustrating a dynamic optimization method based on heterogeneous network topology in one embodiment of the present invention; Figure 2 is a structural diagram illustrating a dynamic optimization device based on heterogeneous network topology in one embodiment of the present invention; Figure 3 is a structural diagram illustrating a dynamic optimization device based on heterogeneous network topology in one embodiment of the present invention; Reference numerals: 11, acquisition module; 12, first processing module; 13, second processing module; 14, modeling module; 15, optimization module; 21, processor; 22, memory. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] An embodiment of the present invention provides a dynamic optimization method based on heterogeneous network topology. Specifically, please refer to Figure 1, which shows a flowchart of the dynamic optimization method based on heterogeneous network topology in one embodiment of the present invention. The method includes: S1: acquiring device status data of the target network; S2: performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; processing the device status data and the device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology graph; S3: inputting the real-time topology graph into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network, wherein... The processing procedure is configured as follows: using the real-time topology graph as a static graph structure, constructing a multimodal feature matrix attached to the static graph structure based on the device nodes and network links of the target network, and using spatiotemporal graph convolutional network technology to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network; S4: acquiring the traffic time series data of the target network, and using generative technology to perform state modeling processing on the traffic time series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network; S5: using multi-agent reinforcement learning technology to perform policy optimization processing on the inference results to obtain a dynamic optimization scheme for the network.

[0022] Specifically, the device status data of the target network is obtained. The device status data includes at least basic identity and configuration status data, hardware health status data, network operation performance status data, and protocol and communication interaction status data.

[0023] The process of performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data includes: performing feature vectorization on the device status data to obtain a high-dimensional feature vector; performing dimensionality reduction and privacy protection processing on the high-dimensional feature vector using a preset sparse random projection matrix to obtain a dimensionality-reduced binary signature; performing equal segmentation on the binary signature; and performing concatenation processing on the obtained signature segments to obtain the device fingerprint data.

[0024] First, the collected multi-source heterogeneous device status data undergoes full-dimensional format normalization and feature extraction, covering all dimensions of status features such as device basic identity, hardware health, network operation performance, protocol communication, fault alarms, etc. Then, features of different data types are uniformly converted into numerical forms. Discrete features are converted into numerical forms using encoding methods, while continuous features are normalized to achieve dimensional uniformity and eliminate feature weight bias caused by differences in numerical ranges. Finally, according to preset feature dimension rules, all the numerically processed features are concatenated in a fixed order to form a high-dimensional feature vector with unified dimensions and complete representation.

[0025] A sparse random projection matrix adapted to the state characteristics of heterogeneous network devices is pre-configured. This matrix is ​​designed with a sparse structure and has the characteristics of low computational complexity. The high-dimensional feature vector obtained in the first step is multiplied with the sparse random projection matrix to complete the projection transformation of the high-dimensional features to the low-dimensional space, resulting in a low-dimensional real-valued vector after projection. Then, a reasonable binarization threshold is set, and each element in the low-dimensional real-valued vector is mapped to a binary form of 0 or 1 with this threshold as the boundary, finally forming a low-dimensional binary signature composed of 0 and 1.

[0026] Next, based on the actual needs of heterogeneous network topology reasoning and device identification, a fixed segment length is preset. The complete binary signature obtained in the second step is divided equally according to this length. If there is a remainder after the division, the last segment is padded according to the preset padding rules to ensure that the length of each segment of the signature is completely consistent. Then, according to the preset fixed order, the several segments of binary signature are concatenated and spliced ​​in sequence to finally form a fixed-length, well-structured binary string or encoded device fingerprint data.

[0027] The device status data and device fingerprint data are processed using incremental maximum likelihood topology reasoning technology to obtain a real-time topology graph. This process includes: performing distributed information exchange processing on the device status data and device fingerprint data to obtain global neighbor relationship tuples and device fingerprint mapping data; processing the global neighbor relationship tuples using incremental maximum likelihood topology reasoning technology to obtain a topology graph structure; and obtaining the real-time topology graph based on the device fingerprint mapping data and the topology graph structure.

[0028] Specifically, relying on the distributed node architecture of a heterogeneous network, all device nodes in the network use device fingerprints as a unified identity identifier. Based on their own collected device status data, such as protocol communication status, port connection status, and neighbor device interaction data, they conduct point-to-point real-time information exchange within the network. The exchange content includes their own device fingerprint, the fingerprints of detected neighbor devices, and the link association status between them, such as link connectivity, communication frequency, and basic link quality. At the same time, through a distributed consensus mechanism, the local neighbor relationship information of each device node is aggregated and integrated across nodes, eliminating false neighbor information caused by communication interference, and forming a global neighbor relationship tuple. The core of the tuple includes three core elements: source device fingerprint, target device fingerprint, and the association attribute of the link between them.

[0029] During the information exchange and aggregation process, device fingerprint mapping data is constructed simultaneously to establish a one-to-one correspondence between device fingerprints and actual basic information of the device, such as unique device identifier, device type, network role, and core hardware status, so as to achieve accurate mapping from fingerprint to device entity.

[0030] Next, incremental information extraction is performed, comparing the global neighbor relationship tuple at the current moment with the tuple at the previous moment to accurately identify incremental changes in the network topology, including newly added device nodes and link associations, disconnected original links, and dynamic changes in link association attributes, such as a decline in communication quality. Inference calculations are performed only on incremental information, rather than re-inferring from all tuples.

[0031] Subsequently, based on the maximum likelihood inference algorithm, probabilistic inference calculations are performed on the global neighbor relation tuple containing incremental information. Based on the actual interaction behavior and link association status between devices, the likelihood probability of a real link association between any two device fingerprints is calculated. A reasonable probability threshold is set to filter out real link associations with a probability higher than the threshold and eliminate false associations with a low probability.

[0032] The inferred valid link associations and device fingerprint nodes are structurally integrated to construct a basic topology graph structure with device fingerprints as nodes and real link associations as edges. This structure includes a set of nodes, a set of links, and basic link association attributes, and can be adapted to an undirected or directed graph structure according to the actual network connection characteristics.

[0033] Next, node entity association is performed. Nodes represented by abstract device fingerprints in the basic topology graph structure are precisely matched using device fingerprint mapping data. The actual device information corresponding to each fingerprint node is attached to the corresponding node in the topology graph, transforming the abstract nodes into entity nodes with actual device attributes. Subsequently, link attribute enrichment is performed. The actual operating attributes corresponding to each link are extracted from the device status data, and these dynamic attributes are associated with the corresponding edges in the topology graph, upgrading the links from simple association relationships to association relationships with operating status. Finally, the topology graph with completed node entityization and link attribute enrichment is structurally regularized and verified, eliminating redundant information and correcting association errors, forming a global real-time topology graph that includes device entity nodes, real link associations between nodes, and the real-time operating status of nodes and links.

[0034] The real-time topology graph is input into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network. The processing is configured to use the real-time topology graph as a static graph structure, construct a multimodal feature matrix attached to the static graph structure based on the device nodes and network links of the target network, and use spatiotemporal graph convolutional network technology to perform feature extraction processing on the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network.

[0035] Using the real-time topology graph as a static graph structure, the multimodal time series data of each vertex and edge within the most recent time window T are used as its features. For a vertex (device node), its feature vector may include: CPU utilization sequence, memory utilization sequence, device temperature sequence, packet transmission / reception rate sequence, etc. For an edge (network link), its feature vector may include: link throughput sequence, latency sequence, jitter sequence, packet loss rate sequence, radio signal strength (RSSI) sequence, etc. These features together constitute a multimodal feature matrix attached to the graph structure.

[0036] Specifically, in the process of using spatiotemporal graph convolutional network technology to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network, the graph convolutional network is used to extract spatial features from the multimodal feature matrix to obtain spatial feature data; the temporal convolutional network is used to extract temporal features from the multimodal feature matrix to obtain temporal feature data; the spatial feature data and the temporal feature data are fused to obtain fused high-order feature data; and the fused high-order feature data is used for root cause localization to obtain the root cause analysis results.

[0037] Specifically, graph convolutional network layers are used to perform message passing and aggregation on the topological graph structure. Each GCN layer aggregates the feature information of the node itself and its first-order neighbor nodes through its learnable weight matrix, thereby effectively capturing the propagation and dependency of network faults in the topological space.

[0038] The constructed multimodal feature matrix is ​​used as the input feature of the graph convolutional network, while the adjacency matrix of the real-time topology graph is used as the spatial correlation constraint of the network and input to the convolutional layer of the graph convolutional network. Through the neighborhood aggregation core operation of graph convolution, the features of each device node / network link are aggregated and weighted with the features of neighboring nodes / links layer by layer. First, the spatial correlation features within the local neighborhood are captured, and then feature fusion within the global topology is achieved through feature propagation of multiple convolutions. This accurately mines the spatial dependencies between devices and links, such as the propagation correlation of faults from core nodes to edge nodes and the cross-domain diffusion correlation of link congestion. Finally, the redundant spatial features are compressed through the pooling layer of the network to extract low-dimensional spatial feature data that integrates local and global topology spatial correlation information.

[0039] It should be noted that the spatial feature data maintains the same node / link dimension as the original feature matrix.

[0040] The time-attributed feature dimensions in the multimodal feature matrix, such as CPU utilization, bandwidth utilization, packet loss rate, and protocol message transmission and reception volume at different acquisition times, are reconstructed into time series. The time series features of the same node / link are organized into a continuous time series feature sequence according to the chronological order to ensure the temporal continuity of the time series features. Then, the time series feature sequence is input into a temporal convolutional network, and through multi-layer convolution operations of causal convolution or one-dimensional convolution, the time series feature sequence is extracted in a sliding window manner to capture the time evolution law of device status and link performance from normal to abnormal, such as the time accumulation trend of hardware load and the gradual aggravation of link congestion.

[0041] Meanwhile, redundant information in the temporal features is removed by pooling layers, key temporal change features are retained, and the extracted temporal features are transformed into temporal feature data that perfectly matches the dimensions of the spatial feature data through feature mapping.

[0042] By adopting a feature-level deep fusion approach, spatial feature data and temporal feature data with unified dimensions are first preliminarily spliced ​​together to form spatiotemporal joint features.

[0043] Then, an attention mechanism is introduced, allowing the model to assign dynamic weights to different dimensions of spatial and temporal features based on the correlation of fault features. High weights are given to feature dimensions that are highly correlated with the root cause of the fault, while low weights are given to irrelevant and redundant features, thus achieving differentiated feature fusion. Subsequently, the weighted spatiotemporal joint features are subjected to feature mapping and dimensionality reduction through a fully connected layer to eliminate dimensional conflicts and information redundancy between features. Finally, a fused high-order feature data that integrates global topological spatial correlation information and the temporal evolution law of device / link state is obtained.

[0044] The root cause localization module of the fault diagnosis model inputs the high-order feature data. First, through the classification sub-module (such as Softmax classifier or multilayer perceptron), based on the feature representation of the high-order features, it accurately determines the fault type of the target network, such as link congestion, equipment failure, protocol conflict, etc., and outputs the fault type identifier.

[0045] Then, through the graph attention localization submodule, combined with the spatial correlation of the real-time topology graph, the fault features in the fused high-order features are traced back to identify the core triggering node / link of the fault (i.e., the root cause node / link of the fault). At the same time, based on the propagation law information in the spatiotemporal fusion features, the cross-space and cross-time propagation path of the fault from the root cause node / link to other nodes / links is derived. Finally, the severity of the fault is quantitatively assessed, and information such as fault type, root cause node / link identification, fault propagation path, and fault severity are integrated to form the root cause analysis results.

[0046] In this embodiment, the fault diagnosis model is trained through unsupervised learning (such as based on reconstruction error: training the model to learn normal patterns, while abnormal data will lead to high reconstruction errors) or semi-supervised learning (training using labeled fault data).

[0047] The model outputs an anomaly score for each node and link. When the anomaly score exceeds a dynamic threshold, the entity is determined to have failed.

[0048] Then, by analyzing the attention weights of the spatial graph convolutional layers in the spatiotemporal graph convolutional network model or calculating the gradient of node features with respect to the final anomaly score (such as using Grad-CAM-like methods), one or a few key entities that contribute the most to the global anomaly can be identified. These key entities are the root causes of the fault.

[0049] Ultimately, the engine outputs structured diagnostic results.

[0050] Generative techniques are used to perform state modeling processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network. This includes: performing synchronous mapping processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain digital twin data; sequentially performing dynamic topology and multimodal state reconstruction processing on the digital twin data to obtain a network state snapshot; and sequentially performing state completion and simulation inference processing on the network state snapshot using generative techniques to obtain the inference results of the target network.

[0051] Synchronization mapping includes time-dimensional synchronization and spatial-dimensional mapping.

[0052] First, time dimension synchronization is carried out, and a unified timestamp granularity is configured for the three types of heterogeneous data. The timestamps are aligned for the continuous time series of traffic time series data, the dynamic update characteristics of device fingerprint data, and the fault time nodes and evolution periods of root cause analysis results. This eliminates the time difference between different data collection and generation stages and achieves accurate matching of the three types of data in the time dimension.

[0053] Secondly, spatial dimension mapping is implemented, using device fingerprints as globally unique association identifiers. The traffic timing characteristics (such as port flow rate and bandwidth usage), static identity characteristics and dynamic status characteristics (such as hardware health and protocol status) of each device / link, and fault characteristics (such as root cause node identifier, fault type, and propagation path) of root cause analysis results are associated one by one with the physical device / link entities corresponding to the target network, thereby achieving the binding of the three types of data in the spatial dimension.

[0054] Finally, feature dimension regularization was completed. Discrete features of the three types of data were uniformly encoded, continuous features were normalized, redundant features were removed and integrated to form a structured feature set. Finally, digital twin data corresponding one-to-one with the physical heterogeneous network was constructed. This data fully carries the network's time, space, device, traffic, and fault information in all dimensions.

[0055] Dynamic topology reconstruction is performed based on the device fingerprint associations in digital twin data and the impact of faults on the topology reflected by root cause analysis results (such as link disconnection caused by the failure of the root cause node and topology logic adjustment caused by link congestion). Combined with the device / link interaction status reflected by traffic time series data, the real-time dynamic topology structure of the target network is restored, replacing the traditional static topology modeling.

[0056] Subsequently, multimodal state reconstruction is carried out. The traffic time-series status of each node / link (such as inbound / outbound flow rate, bandwidth utilization, packet loss rate), the hardware and operating status corresponding to the device fingerprint (such as CPU utilization, protocol interaction status), and the fault status corresponding to the root cause analysis (such as fault severity, fault propagation range, anomaly marker) are extracted from the digital twin data. These multimodal state features are deeply integrated with the reconstructed dynamic topology. According to the preset projection time step, the dynamic topology of the entire network at a certain moment, the multimodal state features of each node / link, and the current development status of the fault are integrated into a complete network state snapshot. The snapshot is time-series structured data, which can completely represent the global operating status of the network at a certain moment.

[0057] During the state completion process, to address the issues of missing device / link state features due to communication interruptions of edge devices and failures of acquisition nodes during physical network data acquisition, as well as abnormal data caused by acquisition noise, a generative model (such as a time-series generation model, a graph generation model, a diffusion model, etc.) is used as the core. Based on the complete feature information in the network state snapshot, the model learns the device association rules, traffic propagation rules, and fault propagation rules of heterogeneous networks to accurately complete the missing node / link state features. At the same time, noise data is corrected and smoothed to ensure the integrity, accuracy, and consistency of each network state snapshot, providing high-quality input data for subsequent simulations.

[0058] Next, a simulation is conducted, using the completed network state snapshot as the initial input. Relying on the temporal feature generation and topology evolution learning capabilities of the generative model, and combining the inherent operating rules of heterogeneous networks (such as the cross-link propagation rules of traffic, the spatial diffusion rate of faults, and the decay rules of equipment performance with load) and the fault characteristics in the root cause analysis results (such as the propagation characteristics of link congestion, the impact range of equipment downtime, and the triggering conditions of protocol conflicts), a preset simulation time range and an uninterrupted simulation scenario are set. Starting from the initial snapshot, the model gradually generates network state snapshots at multiple subsequent time points, simulating the dynamic changes in the topology of the target network, the temporal evolution of the global distribution of traffic, the cross-spatial propagation of faults, and the emergence of potential abnormal nodes / links under the continuous influence of the root cause of the fault and without human intervention.

[0059] Global feature extraction and structured integration are performed on all time-series simulation snapshots generated by the model to extract core information such as the simulation time range, network topology status at each time node, traffic distribution change trend, fault development and propagation path, potential network performance degradation areas, and possible abnormal node / link identifiers in the future, forming standardized and interpretable target network simulation results.

[0060] In this process, the generative technology processing can be summarized as follows: 1. Historical learning and pattern extraction: using generative models such as variational autoencoders (VAE) or generative adversarial networks (GAN), the joint probability distribution of network multimodal data under historical normal state and various fault states is learned, and its dynamic evolution pattern is mastered.

[0061] 2. Real-time state completion: When some entity data is missing, the pre-trained generative model is used to generate the most likely missing state data based on the observed state of surrounding entities, thus completing the construction of the twin's panoramic state.

[0062] 3. Short-term state prediction: Using time-series prediction models (such as LSTM, TCN, or Transformer) as input, predict the network state (such as traffic matrix, load changes) in the next few time slices, providing support for forward-looking decision-making.

[0063] The network dynamic optimization scheme is obtained by using multi-agent reinforcement learning technology to optimize the inference results and obtain a joint policy model.

[0064] In this process, a multi-agent architecture needs to be deployed. First, based on the topology of the target heterogeneous network, the networking roles of devices (core forwarding nodes, edge access nodes, fault handling nodes, etc.) and the network domain division characteristics in the inference results, a distributed agent cluster corresponding one-to-one with physical network nodes / domains is constructed. The device fingerprint is used as the unique identifier of the agent, so that each agent can accurately correspond to the actual device / domain in the network. At the same time, distributed communication rules between agents are configured to ensure collaborative interaction without protocol differences.

[0065] Secondly, a standardized learning space is defined. Core information extracted from the simulation results, such as network topology status, global traffic distribution trends, fault propagation patterns, potential risk node / link attributes, and network performance index thresholds, is transformed into an environmental state space that can be uniformly perceived by multiple agents. At the same time, combined with the operation and maintenance types of heterogeneous networks, the executable action space of each agent is set, such as dynamic bandwidth adjustment, route replanning, fault node isolation, cross-link traffic splitting, and device load balancing. A global collaborative reward function is designed, with positive rewards for overall network performance improvement (reduced latency, reduced packet loss rate), fault propagation suppression efficiency, and optimized resource utilization, and negative penalties for network oscillations caused by operations, delayed fault handling, and resource waste. This abandons the local reward logic of single agents and ensures that policy learning is directed towards global optimization.

[0066] Subsequently, multi-agent collaborative training was conducted. Multi-agent reinforcement learning algorithms adapted to distributed networks (such as MADDPG and QMIX) were selected, and the inference results were used as the core training samples to simulate a network evolution environment without human intervention. Each agent, based on its own perceived local state, such as the load of its node, the traffic of neighboring links, and collaborative interaction information between agents, continuously tried and adjusted its actions and received reward feedback. It gradually learned each other's action decision logic, solved the information asymmetry problem of distributed agents, and avoided local optima. Finally, when the training process reached the convergence threshold, that is, the global reward was stable and the network performance optimization effect met the target, the action decision logic of each agent, the collaborative interaction rules between agents, and the mapping relationship between global state and action were solidified to form a joint policy model that can be directly invoked. This model has the ability to make autonomous decisions and collaborative optimizations to adapt to the dynamic changes of the network.

[0067] In the process of generating the optimal policy for the joint policy model, state fusion correction is first performed. The real-time actual operating state of the target network is compared and fused with the inference results to correct the deviation between the inference results and the actual state. The corrected global state features are then transformed into standardized state data that meet the input requirements of the joint policy model, thus avoiding policy inaccuracies caused by prediction deviations from pure inference.

[0068] Subsequently, model invocation and candidate action generation are performed. The corrected standardized state data is input into the joint policy model. Based on the fixed collaborative decision-making logic, the model drives each distributed agent to independently output a set of candidate actions adapted to itself based on its corresponding network node / domain state. At the same time, through the collaborative rules in the model, the initial interaction and matching of candidate actions among the agents are realized.

[0069] Next, global action screening and collaborative verification are carried out. Candidate actions of all agents are verified in two layers: first, feasibility verification, which eliminates actions that cannot be actually executed by combining device hardware performance, link bandwidth limit, and network protocol constraints; second, conflict detection and resolution, which identifies and eliminates action conflicts between different agents (such as two agents adjusting the bandwidth of the same link at the same time, or path conflicts caused by multiple route replanning), and performs global combination optimization on conflict-free and executable actions to ensure that the action combination can achieve the best global network performance.

[0070] Then, the strategy is transformed into operational implementation. The optimized global collaborative action set is converted into standardized operation and maintenance operation instructions that can be directly executed by heterogeneous network devices. Using device fingerprints as identifiers, the operation instructions are accurately mapped to specific devices / links. The instructions clearly include core information such as the operation object (fingerprint identifier), specific operation type, operation parameters (such as bandwidth adjustment ratio, traffic diversion coefficient), and operation execution conditions. Finally, the solution is structurally integrated. All operation and maintenance operation instructions are classified and organized according to the handling priority (fault root cause node handling first, core link traffic scheduling first), execution sequence (isolate fault nodes first and then perform traffic diversion, re-plan routes first and then adjust bandwidth), and the network domain to which the operation belongs. At the same time, the expected effect of each operation, the network performance optimization target after execution, and the acceptance criteria for fault handling are clearly defined, ultimately forming a structured, implementable, and step-by-step network dynamic optimization solution.

[0071] After obtaining the network dynamic optimization scheme, the dynamic optimization method based on heterogeneous network topology further includes: executing the network dynamic optimization scheme to obtain network status feedback data; performing difference comparison processing on the network status feedback data to obtain a simulation difference report; and performing cognitive model optimization processing on the simulation difference report to obtain the optimized network dynamic optimization scheme.

[0072] Another embodiment of the present invention provides a dynamic optimization device based on heterogeneous network topology. Specifically, please refer to Figure 2, which shows a schematic diagram of the structure of a dynamic optimization device based on heterogeneous network topology in one embodiment of the present invention. The device includes: an acquisition module 11, used to acquire device status data of the target network; a first processing module 12, used to perform protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; and to process the device status data and the device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology graph; and a second processing module 13, used to input the real-time topology graph into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network. The processing procedure is configured to use the real-time topology graph as a static graph structure, construct a multimodal feature matrix attached to the static graph structure based on the device nodes and network links of the target network, and use spatiotemporal graph convolutional network technology to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network; modeling module 14 is used to acquire the traffic time series data of the target network, and use generative technology to perform state modeling processing on the traffic time series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network; optimization module 15 is used to use multi-agent reinforcement learning technology to perform policy optimization processing on the inference results to obtain a dynamic optimization scheme for the network.

[0073] Referring to Figure 3, which is a schematic diagram of the structure of the dynamic optimization device based on heterogeneous network topology provided in an embodiment of the present invention, the dynamic optimization device based on heterogeneous network topology provided in an embodiment of the present invention includes a processor 21, a memory 22, and a computer program stored in the memory 22 and configured to be executed by the processor 21. When the processor 21 executes the computer program, it implements the steps as described in the above embodiments of the dynamic optimization method based on heterogeneous network topology, such as steps S1 to S5 in Figure 1; or, when the processor 21 executes the computer program, it implements the functions of each module in the above embodiments of the device, such as the acquisition module 11.

[0074] For example, the computer program can be divided into one or more modules, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the dynamic optimization device based on heterogeneous network topology. For example, the computer program can be divided into an acquisition module 11, a first processing module 12, a second processing module 13, etc., with the specific functions of each module as follows: the acquisition module 11 is used to acquire device status data of the target network; the first processing module 12 is used to perform protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; the device status data and the device fingerprint data are processed using incremental maximum likelihood topology inference technology to obtain a real-time topology map; the second processing module 13 is used to input the real-time topology map into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network, wherein the processing is configured to... The real-time topology graph is a static graph structure. A multimodal feature matrix is ​​constructed on the static graph structure based on the device nodes and network links of the target network. The spatiotemporal graph convolutional network technology is used to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network. The modeling module 14 is used to acquire the traffic time series data of the target network and use generative technology to perform state modeling processing on the traffic time series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network. The optimization module 15 is used to use multi-agent reinforcement learning technology to perform policy optimization processing on the inference results to obtain a dynamic optimization scheme for the network.

[0075] The dynamic optimization device based on heterogeneous network topology may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will understand that the schematic diagram is merely an example of a dynamic optimization device based on heterogeneous network topology and does not constitute a limitation on the device. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the dynamic optimization device based on heterogeneous network topology may also include input / output devices, network access devices, buses, etc.

[0076] The processor 21 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 21 is the control center of the dynamic optimization device based on heterogeneous network topology, connecting various parts of the device through various interfaces and lines.

[0077] The memory 22 can be used to store the computer programs and / or modules. The processor 21 implements various functions of the dynamic optimization device based on heterogeneous network topology by running or executing the computer programs and / or modules stored in the memory 22 and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0078] Wherein, if the module integrated by the dynamic optimization device based on heterogeneous network topology is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0079] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0080] Accordingly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform steps in the dynamic optimization method based on heterogeneous network topology as described in the above embodiments, such as steps S1 to S5 as shown in FIG1.

[0081] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A dynamic optimization method based on heterogeneous network topology, characterized in that, include: The process involves: acquiring device status data of the target network; performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; processing the device status data and device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology graph; inputting the real-time topology graph into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing to obtain the root cause analysis results of the target network; wherein the processing is configured to use the real-time topology graph as a static graph structure, constructing a multimodal feature matrix attached to the static graph structure based on the device nodes and network links of the target network, and using spatiotemporal graph convolutional network technology to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network; acquiring traffic time-series data of the target network, and using generative techniques to perform state modeling processing on the traffic time-series data, device fingerprint data, and root cause analysis results to obtain the inference results of the target network; and using multi-agent reinforcement learning technology to perform policy optimization processing on the inference results to obtain a dynamic network optimization scheme.

2. The dynamic optimization method based on heterogeneous network topology as described in claim 1, characterized in that, The step of performing protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data includes: performing feature vectorization on the device status data to obtain a high-dimensional feature vector; performing dimensionality reduction and privacy protection processing on the high-dimensional feature vector using a preset sparse random projection matrix to obtain a dimensionality-reduced binary signature; performing equal segmentation on the binary signature; and performing concatenation processing based on the obtained several signature segments to obtain the device fingerprint data.

3. The dynamic optimization method based on heterogeneous network topology as described in claim 1, characterized in that, The step of processing the device status data and device fingerprint data using incremental maximum likelihood topological reasoning technology to obtain a real-time topological graph includes: performing distributed information exchange processing on the device status data and device fingerprint data to obtain global neighbor relation tuples and device fingerprint mapping data; processing the global neighbor relation tuples using incremental maximum likelihood topological reasoning technology to obtain a topological graph structure; and obtaining the real-time topological graph based on the device fingerprint mapping data and the topological graph structure.

4. The dynamic optimization method based on heterogeneous network topology as described in claim 1, characterized in that, The step of using spatiotemporal graph convolutional network technology to extract spatial and temporal features from the multimodal feature matrix to obtain the root cause analysis results of the target network includes: using a graph convolutional network to extract spatial features from the multimodal feature matrix to obtain spatial feature data; using a temporal convolutional network to extract temporal features from the multimodal feature matrix to obtain temporal feature data; fusing the spatial feature data and the temporal feature data to obtain fused high-order feature data; and performing root cause localization processing on the fused high-order feature data to obtain the root cause analysis results.

5. The dynamic optimization method based on heterogeneous network topology as described in claim 1, characterized in that, The process of using generative techniques to perform state modeling on the traffic time-series data, device fingerprint data, and root cause analysis results to obtain the inference results of the target network includes: performing synchronous mapping processing on the traffic time-series data, device fingerprint data, and root cause analysis results to obtain digital twin data; sequentially performing dynamic topology and multimodal state reconstruction processing on the digital twin data to obtain a network state snapshot; and sequentially performing state completion and simulation inference processing on the network state snapshot using generative techniques to obtain the inference results of the target network.

6. The dynamic optimization method based on heterogeneous network topology as described in claim 1, characterized in that, The step of using multi-agent reinforcement learning technology to optimize the inference results and obtain a dynamic network optimization scheme includes: using multi-agent reinforcement learning technology to optimize the inference results and obtain a joint policy model; and performing optimal policy generation processing on the joint policy model to obtain the dynamic network optimization scheme.

7. The dynamic optimization method based on heterogeneous network topology as described in claim 1, characterized in that, After obtaining the network dynamic optimization scheme, the dynamic optimization method based on heterogeneous network topology further includes: executing the network dynamic optimization scheme to obtain network status feedback data; performing difference comparison processing on the network status feedback data to obtain a simulation difference report; and performing cognitive model optimization processing on the simulation difference report to obtain the optimized network dynamic optimization scheme.

8. A dynamic optimization device based on heterogeneous network topology, characterized in that, include: The acquisition module is used to acquire device status data of the target network. The first processing module is used to perform protocol-insensitive sparse feature hashing on the device status data to obtain device fingerprint data; and to process the device status data and the device fingerprint data using incremental maximum likelihood topology inference technology to obtain a real-time topology map. The second processing module is used to input the real-time topology graph into a fault diagnosis model constructed by a spatiotemporal graph convolutional network for processing, to obtain the root cause analysis results of the target network. The processing is configured to use the real-time topology graph as a static graph structure, construct a multimodal feature matrix attached to the static graph structure based on the device nodes and network links of the target network, and use spatiotemporal graph convolutional network technology to extract features from the multimodal feature matrix in both spatial and temporal dimensions to obtain the root cause analysis results of the target network. The modeling module is used to acquire the traffic time-series data of the target network, and use generative techniques to perform state modeling processing on the traffic time-series data, the device fingerprint data, and the root cause analysis results to obtain the inference results of the target network. The optimization module is used to use multi-agent reinforcement learning technology to perform policy optimization processing on the inference results to obtain a dynamic network optimization scheme.

9. A dynamic optimization device based on heterogeneous network topology, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the dynamic optimization method based on heterogeneous network topology as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the dynamic optimization method based on heterogeneous network topology as described in any one of claims 1 to 7.

Images