Method and device for determining health state of video monitoring network and computer equipment

Abstract

The invention relates to a health state determination method and device of a video monitoring network and computer equipment, and the method comprises the steps: obtaining multi-source video monitoring data of the video monitoring network, the multi-source video monitoring data comprising equipment and operation state data, network flow data, video quality data and user operation log data; based on a pre-constructed network state semantic map, performing semantic fusion on the multi-source video monitoring data to obtain semantic association features; based on the semantic association features, performing health degree evaluation by using a pre-trained health degree evaluation model to obtain a health state of the video monitoring network; wherein the health degree evaluation model is obtained through network architecture training of a fusion graph neural network and a long short-term memory network; and under the condition that the health state represents that the health of the video monitoring network is abnormal, an abnormal traceability result is determined through the health degree evaluation model, so that the comprehensive and accurate perception capability of health degree evaluation on the network state can be improved.

Description

Technical Field

[0001] This application relates to the field of network health assessment technology, and in particular to a method, apparatus and computer equipment for determining the health status of a video surveillance network. Background Technology

[0002] In current urban security, traffic management, and other application scenarios, the deployment scale and complexity of video surveillance networks continue to increase. Traditional video surveillance systems have formed a basic health assessment system by collecting multi-dimensional data such as equipment operating status, network traffic, and video quality, providing basic support for system operation and maintenance.

[0003] With the increasing complexity of network structures and the continuous expansion of application scenarios, existing monitoring systems face new challenges in handling the deep integration and intelligent analysis of multi-source heterogeneous data. In particular, when dealing with complex scenarios such as spatial dependencies between network devices and the dynamic evolution of anomaly patterns, the adaptability and accuracy of existing evaluation methods urgently need to be improved.

[0004] The existing health assessment system for video surveillance networks has significant limitations in terms of data fusion depth and spatiotemporal correlation modeling, making it difficult to achieve comprehensive and accurate perception and intelligent intervention of network operation status. This has become a key technical bottleneck restricting the improvement of the operation and maintenance efficiency of video surveillance networks. Summary of the Invention

[0005] Based on this, the purpose of this application is to at least solve one of the above-mentioned technical defects, especially the technical defect that it is difficult to achieve a comprehensive and accurate perception of network status in the prior art. This application provides a method, apparatus and computer equipment for determining the health status of a video surveillance network.

[0006] Firstly, this application provides a method for determining the health status of a video surveillance network, the method comprising:

[0007] Acquire multi-source video surveillance data from the video surveillance network. This multi-source video surveillance data includes equipment and operating status data, network traffic data, video quality data, and user operation log data.

[0008] Based on a pre-constructed network state semantic graph, semantic fusion is performed on multi-source video surveillance data to obtain semantic association features;

[0009] Based on semantic association features, a pre-trained health assessment model is used to assess the health status of the video surveillance network. The health assessment model is trained using a network architecture that integrates graph neural networks and long short-term memory networks.

[0010] When the health status characterizes an abnormality in the video surveillance network, the source of the abnormality can be determined through a health assessment model.

[0011] In one embodiment, before semantically fusing multi-source video surveillance data based on a pre-built network state semantic graph to obtain semantic association features, the process includes:

[0012] Construct a semantic graph for video surveillance networks, using devices, links, fault types, and video quality indicators as entities, and establish entity relationships between entities through relation edges;

[0013] Based on the various entities and their relationships, a network state semantic graph is constructed.

[0014] In one embodiment, semantic fusion of multi-source video surveillance data is performed based on a pre-constructed network state semantic graph to obtain semantic association features, including:

[0015] Determine the weight parameters corresponding to multi-source video surveillance data;

[0016] Based on the entity relationships and weight parameters of the network state semantic graph, multi-source video surveillance data are weighted and fused to obtain semantic association features.

[0017] In one embodiment, based on semantic association features, a pre-trained health assessment model is used to assess the health status of the video surveillance network, including:

[0018] The semantic association features are input into the health assessment model, the spatial dependencies between entities are learned through graph neural networks, and the temporal change features of the network state are captured through long short-term memory networks.

[0019] By integrating spatial dependence and temporal variation characteristics, a health score is output, which represents the health status.

[0020] In one embodiment, the anomaly tracing results are determined using a health assessment model, including:

[0021] The anomaly detection submodule built into the health assessment model is used to locate abnormal data features based on the health score.

[0022] By tracing the root cause of anomalies based on the characteristics of abnormal data and the semantic graph of network state, the anomaly tracing results are obtained.

[0023] In one embodiment, the method further includes:

[0024] When the health status representation network has a pre-defined mild abnormality, an optimized intervention action is selected based on a pre-defined reinforcement learning algorithm.

[0025] Simulate and optimize intervention actions in a digital twin environment, and quantify the changes in health status before and after the intervention using a health assessment model;

[0026] When the results of changes in health status indicate that the video surveillance network is in good health, the optimized intervention actions are applied to the video surveillance network to improve its health.

[0027] In one embodiment, the method further includes:

[0028] Based on entity relationships in network state semantic graphs and multi-source video surveillance data, scene types are identified;

[0029] Adjust the model parameters of the health assessment model and the health threshold used for health status assessment according to the scenario type to improve the universality of video surveillance networks of different scales and environments.

[0030] Secondly, this application provides a health status determination device for a video surveillance network, the device comprising:

[0031] The data acquisition module is used to acquire multi-source video surveillance data from the video surveillance network. The multi-source video surveillance data includes equipment and operating status data, network traffic data, video quality data, and user operation log data.

[0032] The semantic fusion module is used to semantically fuse multi-source video surveillance data based on a pre-built network state semantic graph to obtain semantic association features;

[0033] The health assessment module is used to assess the health status of the video surveillance network based on semantic association features and a pre-trained health assessment model. The health assessment model is trained using a network architecture that integrates graph neural networks and long short-term memory networks.

[0034] The anomaly tracing module is used to determine the anomaly tracing results by using a health assessment model when the health status characterizes an anomaly in the video surveillance network.

[0035] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0036] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0037] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0038] This application provides a method, apparatus, and computer equipment for determining the health status of a video surveillance network.

[0039] By acquiring multi-source video surveillance data, including equipment and operational status data, network traffic data, video quality data, and user operation log data, a comprehensive overview of the video surveillance network's equipment operation, network transmission, video content quality, and user interaction can be achieved. This avoids the limitation of a single data dimension failing to fully reflect the overall health of the network. Furthermore, by semantically fusing multi-source video surveillance data through a pre-constructed network status semantic graph, a health assessment model can be used based on semantic association features to obtain the health status of the video surveillance network. This allows for the processing of dispersed structured / unstructured data. This data is transformed into a feature set with semantic and logical connections, allowing for the discovery of hidden relationships between multi-source data (such as the correlation between abnormal network traffic and video stuttering, or the correlation between frequent user operations and excessive device load). This improves the accuracy of health assessment in identifying network status. If the health status indicates an abnormality in the video surveillance network, the health assessment model can also determine the source of the anomaly. In this way, the root cause of the anomaly can be quickly located (such as device hardware failure, network link congestion, abnormal video encoding algorithm, or user misoperation), reducing the time cost of troubleshooting, improving the efficiency of anomaly repair, and reducing the impact of video surveillance network anomalies on the continuity of monitoring services. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a method for determining the health status of a video surveillance network, provided in an embodiment of this application;

[0042] Figure 2 A schematic diagram of an architecture for semantic fusion of multi-source video surveillance data provided in an embodiment of this application;

[0043] Figure 3 A flowchart illustrating a step for determining health status, provided as an embodiment of this application;

[0044] Figure 4 This is a schematic diagram of the structure of a GNN-LSTM hybrid model provided in an embodiment of this application;

[0045] Figure 5A flowchart illustrating a reinforcement learning-driven active intervention strategy optimization method provided in an embodiment of this application;

[0046] Figure 6 A schematic diagram of a scenario-based dynamic adaptation architecture provided for an embodiment of this application;

[0047] Figure 7 This is a schematic diagram of the structure of a health status determination device for a video surveillance network provided in an embodiment of this application;

[0048] Figure 8 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0050] With the widespread deployment of video surveillance networks in urban security, traffic management, and corporate parks, their scale and complexity are constantly increasing. Traditional video surveillance network monitoring methods have many limitations:

[0051] 1. The data fusion methods are too simplistic, mostly involving only simple data splicing, failing to uncover the deep semantic relationships between multiple data sources. This results in insufficient support for network health assessment after fusion. For example, when faced with concurrent network congestion and equipment failures, it is impossible to accurately determine the combined impact of both on network health.

[0052] 2. Health assessment models lack collaborative modeling of network spatial topology and time series. Existing models mostly focus on time series analysis, neglecting the spatial dependencies between network nodes. For example, if multiple cameras in a certain area experience video transmission abnormalities due to the failure of the same switch, existing models cannot fully capture such complex situations.

[0053] 3. In the proactive intervention phase, there is a heavy reliance on preset rules, lacking adaptability and intelligent evolutionary capabilities. When new abnormal patterns appear in the network, the intervention strategy cannot be adjusted quickly, resulting in poor intervention effects and potentially triggering secondary failures.

[0054] Meanwhile, while video surveillance networks possess abundant multi-source data, including equipment operation status data, network traffic data, video quality data, and user operation logs, this data is scattered and underutilized. How to effectively integrate this data and build a comprehensive, intelligent health assessment and proactive intervention system has become an urgent problem to be solved.

[0055] In one exemplary embodiment, Figure 1 This is a flowchart illustrating a method for determining the health status of a video surveillance network, as provided in an embodiment of this application. Figure 1 As shown, a method for determining the health status of a video surveillance network is provided. This example illustrates the method applied to a server; however, it can also be applied to terminals and systems including both terminals and servers, and is implemented through interaction between the terminals and servers. Terminals can be, but are not limited to, various personal computers, laptops, smartphones, and tablets. Servers can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing cloud computing services. In this embodiment, the method includes the following steps S101 to S104: Wherein:

[0056] S101. Obtain multi-source video surveillance data from the video surveillance network. The multi-source video surveillance data includes equipment and operating status data, network traffic data, video quality data, and user operation log data.

[0057] Multi-source video surveillance data refers to heterogeneous data sets from different stages and dimensions within the video surveillance network. Equipment and operational status data can include the online / offline status of hardware such as cameras, encoders, and storage devices, CPU / memory utilization, temperature, and fault alarm information. Network traffic data describes the bandwidth usage, packet loss rate, latency, and jitter of the monitoring data stream during transmission. Video quality data quantifies the clarity, smoothness, and color fidelity of the video signal itself, including indicators such as signal strength, pixelation, and number of stutters. User operation log data records the sequence of actions performed by maintenance personnel, such as logging in, modifying configurations, and accessing and playing back videos.

[0058] For example, the server can periodically or event-triggeredly collect the above four types of data from network devices, video platforms, and security systems by calling the SNMP protocol, NetFlow flow statistics technology, video quality diagnostic algorithms, and system operation log interfaces provided by the device manufacturer, and perform preliminary cleaning and formatting to provide a unified data foundation for subsequent fusion analysis.

[0059] S102. Based on the pre-constructed network state semantic graph, semantic fusion is performed on multi-source video surveillance data to obtain semantic association features.

[0060] The network state semantic graph refers to a knowledge base describing the elements and complex relationships within a video surveillance network. It can be organized in a graph structure, with nodes representing entities and edges representing relationships between entities. It can also be a graph-based data structure composed of nodes and edges, used in this application to construct semantic relationships between entities such as devices, links, fault types, and video quality indicators in the video surveillance network. Semantic fusion refers to the process of integrating raw, heterogeneous multi-source data into a unified semantic feature vector based on their logical relationships within the semantic graph. Semantic relationship features refer to the fused, structured feature representation that characterizes the overall network operation and reveals the underlying causal and influence chains behind the data.

[0061] For example, after acquiring multi-source video surveillance data, the server can align the multi-source video surveillance data with a pre-constructed network state semantic graph. For instance, the device state data "camera A's CPU is too high" can be associated with the entity "network switch port connected to camera A" and the relation edge "cause" to form a semantic association chain of "abnormal device load  may cause  abnormal network transmission  thus affecting  video quality degradation". This chain can be quantized into a feature vector to obtain semantic association features.

[0062] S103. Based on semantic association features, a pre-trained health assessment model is used to assess the health status of the video surveillance network and obtain the health status of the network. The health assessment model is trained using a network architecture that integrates graph neural networks and long short-term memory networks.

[0063] Among these, the health assessment model refers to a machine learning model capable of comprehensively evaluating the overall operational status of a video surveillance network. Health status can be a comprehensive quantitative indicator or level label, used to intuitively reflect the continuous state spectrum of the network from "healthy" to "faulty." Graph Neural Networks (GNNs) are deep learning models specifically designed for processing graph-structured data, adept at learning spatial dependencies between nodes in a graph. For example, GNNs can specifically handle graph-structured data, learning the relationships and dependencies between nodes in a graph, and are used to capture the spatial topological relationships of a video surveillance network. Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network, adept at capturing long-term temporal variation features from time-series data. For example, LSTMs are suitable for processing and predicting time-series data, and can capture the changing trends of the video surveillance network's status over time.

[0064] For example, the server can input the semantic association feature sequence obtained in S102 into the health assessment model. The health assessment model can use graph neural networks to analyze the mutual influence relationships (spatial dependencies) between various entities (devices, links, etc.) in the network at the current moment; use long short-term memory networks to analyze the state change trends (temporal changes) of these entities and relationships within historical time windows; and can fuse features from both spatial topology and temporal data to output a comprehensive health score or state level.

[0065] S104. When the health status characterizes an abnormality in the video surveillance network, the source of the abnormality is determined through the health assessment model.

[0066] The anomaly tracing results refer to a detailed report that identifies the root cause of the anomaly and its impact path after determining that the network health is abnormal. The anomaly tracing results can indicate which component is experiencing a problem and explain how the problem spreads to other parts of the network through its interconnections, ultimately leading to a decline in overall health.

[0067] For example, when the health assessment result of the health assessment model is lower than a preset threshold, a source tracing analysis can be triggered. The source tracing module built into the health assessment model can trace back its decision path during the assessment process, combine it with the network state semantic graph, identify the abnormal data features that contribute the most to the decline in health (e.g., a persistently high packet loss rate of a core switch), and deduce the chain reaction that the anomaly may cause by following the relationship edges in the graph, ultimately generating a source tracing report that includes the root cause location, the scope of impact, and possible causes.

[0068] In this embodiment, by acquiring and fusing multi-dimensional data such as device, network, quality, and operation logs, a full-chain perception system covering the video surveillance network is constructed, overcoming the limitations of a single data source perspective. By utilizing a network state semantic graph to semantically correlate multi-source heterogeneous data, isolated alarms and indicators are transformed into logically related features, achieving comprehensive and accurate perception of network operating status from isolated points to a network of relationships. Furthermore, through a deep learning model that integrates spatiotemporal features for comprehensive evaluation and intelligent source tracing, health status judgment is no longer a simple threshold comparison but a deep analysis based on the inherent relationships within the network, thereby improving the accuracy of perceiving the operating status of the video surveillance network.

[0069] In an exemplary embodiment, before semantically fusing multi-source video surveillance data based on a pre-built network state semantic graph to obtain semantic association features, the process includes:

[0070] Construct a semantic graph for video surveillance networks, using devices, links, fault types, and video quality indicators as entities, and establish entity relationships between entities through relation edges;

[0071] Based on the various entities and their relationships, a network state semantic graph is constructed.

[0072] In this context, "entity" refers to a description of a specific thing or abstract concept in a knowledge graph, and in this application, it refers to a core component of a video surveillance network. Device entities can include cameras, NVRs, switches, servers, etc. Link entities can describe the physical or logical connections between devices. Fault type entities can provide standardized definitions for various known faults. Video quality metric entities can cover quantitative parameters such as resolution, frame rate, and latency. Entity relationships refer to directed edges connecting two or more entities, used to describe the interactions between entities, such as "located in," "connected," "cause," and "affect."

[0073] For example, the server can automatically or semi-automatically extract entities and relationships from network topology diagrams, device configuration libraries, historical fault databases, and industry standard documents. For instance, if it is identified that "camera C" is connected to "switch S" via "network cable L", a triple (camera C, connection, switch S) can be created. At the same time, based on operational experience, rules such as "high load on switch S" "causes" "video stream stuttering" can be defined.

[0074] In some exemplary implementations, a knowledge graph-driven association network can be constructed: a network state semantic graph for video surveillance networks is constructed, with devices, links, fault types, video quality indicators, etc., as entities, and semantic associations are established through relational edges.

[0075] In this embodiment, by constructing a structured network state semantic graph, the originally loose and heterogeneous network data is unified into a single knowledge network, providing a semantic foundation for the deep fusion of multi-source data. This enables the understanding of the business logic and causal relationships behind the data, laying the foundation for achieving a leap from "data piling up" to "knowledge association" and realizing a comprehensive and accurate perception of the network's operational status.

[0076] In one exemplary embodiment, semantic fusion of multi-source video surveillance data is performed based on a pre-constructed network state semantic graph to obtain semantic association features, including:

[0077] Determine the weight parameters corresponding to multi-source video surveillance data;

[0078] Based on the entity relationships and weight parameters of the network state semantic graph, multi-source video surveillance data are weighted and fused to obtain semantic association features.

[0079] The weighting parameter refers to a coefficient used to measure the importance of different monitoring data in the current context. The weighting parameter can be dynamically adjusted to reflect the differences in the contribution of different data sources to the network health status. Weighted fusion refers to a data integration strategy that uses weighted parameters to sum or concatenate data from different sources and of different types to highlight key information and suppress noise interference.

[0080] For example, the server can adaptively calculate weights based on the freshness of multi-source video surveillance data, the reliability of the data source, and its relevance to the current scenario of interest. For instance, when investigating video stuttering issues, the weights of network traffic data and video quality data might be increased, while the weight of user operation logs might be relatively decreased. Based on the association paths defined in the semantic graph, high-weighted abnormal data (such as high packet loss rates) can be bound to their associated entities (such as specific links and devices), generating semantic association features that better reflect the essence of the problem.

[0081] In this embodiment, by introducing dynamic weight parameters for weighted fusion, the fused semantic association features can more accurately focus on the factors that have the most significant impact on network health, thereby improving the sensitivity and accuracy of state perception.

[0082] In some exemplary implementations, data features can be weighted based on entity relationships during data fusion. For example... Figure 2 As shown, Figure 2 This application provides a schematic diagram of an architecture for semantic fusion of multi-source video surveillance data, wherein:

[0083] Multi-source data acquisition includes: "device operating status data (CPU / memory, etc.)", "network traffic data (bandwidth / packet loss rate, etc.)", "video quality data (frame rate / resolution, etc.)" and "user operation log data (login / operation records, etc.)".

[0084] Network state semantic graph construction: such as "camera-connection-switch" and "high packet loss rate-cause-video stuttering".

[0085] Feature extraction and weighting: For example, the weight of switch data is increased by 30%.

[0086] Dynamic weight adjustment: such as "scene labels (e.g. peak traffic / off-peak traffic)," based on historical data, learns the scene to generate weight mapping relationship, adjusts the parameters of the feature weighting process, and outputs "fusion feature vector."

[0087] By integrating feature extraction and weighting of video surveillance data, entity association weight rules, and dynamic weight adjustment, feature weights can be assigned according to the strength of association.

[0088] Based on historical data, the learning scenario-weight mapping relationship can be used to achieve semantic data fusion, which can then be used as input for subsequent health assessments.

[0089] In one exemplary embodiment, such as Figure 3 As shown, Figure 3 This application provides a flowchart illustrating a step for determining health status, which can be implemented in an embodiment of the present application. Figure 1 Based on this, the steps of the method for determining the health status of a video surveillance network are illustrated by example. In step S103, based on semantic association features, a pre-trained health assessment model is used to perform a health assessment to obtain the health status of the video surveillance network, including:

[0090] S301. Input semantic association features into the health assessment model, learn the spatial dependency relationship between entities through graph neural network, and capture the temporal change features of network state through long short-term memory network.

[0091] S302. Integrate spatial dependency and temporal variation characteristics to output a health score, which represents the health status.

[0092] Spatial dependency refers to the mutual influence and constraint relationships between various entities (such as cameras and switches) in a video surveillance network due to their physical connections or logical associations. Temporal variation characteristics refer to the dynamic changes, trends, and periodic patterns exhibited by network state parameters (such as traffic, load, and quality) over time. Fusion refers to the effective combination and interaction of different types (spatial and temporal) features to form a more comprehensive and robust feature representation.

[0093] For example, the GNN component processes semantic association features organized in a graph structure. Through a message passing mechanism, each entity node aggregates information from its neighboring nodes, thereby learning the local and global spatial structure of the network. The LSTM component receives the temporally ordered feature sequences of each entity or the entire network. Through its internal gating mechanism, it memorizes long-term dependencies and captures temporal patterns such as "slow load increases" or "periodic peaks." Finally, the model concatenates and transforms the spatial feature vector output by the GNN and the temporal feature vector output by the LSTM through fully connected layers or attention mechanisms, mapping them to a final health score.

[0094] In this embodiment, by fusing graph neural networks and long short-term memory networks, the health assessment model simultaneously possesses the ability to perceive both the "spatial structural complexity" and the "temporal dynamic evolution" of the network. This spatiotemporal joint analysis can more comprehensively perceive the network status. For example, it can not only detect "high load on a certain switch" (spatial), but also determine "whether it is a sudden situation or a continuous deterioration" (temporal), thereby achieving a deep, accurate, and forward-looking perception of the network health status and avoiding misjudgments or omissions caused by ignoring either spatiotemporal dimension.

[0095] In some exemplary embodiments, such as Figure 4 As shown, Figure 4 A schematic diagram of the structure of a GNN-LSTM hybrid model provided in this application embodiment, wherein:

[0096] Input layer: "Fusion feature vector (including spatial / temporal features)", which includes three sub-items: "Device node features", "Link connection features", and "Temporal variation features".

[0097] GNN (Fusion Graph Neural Network) module: The left side is "Graph Neural Network (Spatial Feature Extraction)", which contains 3 node layers (the number of nodes in each layer is 64, 32 and 16 respectively) and outputs "spatial feature vectors (such as node dependencies)".

[0098] LSTM (Long Short-Term Memory) module: The upper neural network structure on the right, "Long Short-Term Memory Network (Time Feature Extraction)," contains two LSTM unit layers (32 units per layer), which are time-series dependent on each other, and output "time feature vectors (such as state change trends)."

[0099] Fusion layer: Feature cross-fusion, receives the outputs of GNN and LSTM, performs feature fusion, and outputs "spatiotemporal comprehensive features".

[0100] Output layer: The top elliptical box “Health Score (0-100)” serves as the final output of the model.

[0101] In one exemplary embodiment, the anomaly tracing results are determined using a health assessment model, including:

[0102] The anomaly detection submodule built into the health assessment model is used to locate abnormal data features based on the health score.

[0103] By tracing the root cause of anomalies based on the characteristics of abnormal data and the semantic graph of network state, the anomaly tracing results are obtained.

[0104] The anomaly detection submodule refers to a component within the health assessment model specifically responsible for identifying and locating abnormal patterns. This can be implemented based on attention mechanisms, gradient analysis, or specialized anomaly detection algorithms. Locating anomalous data features refers to accurately identifying which data features (and their corresponding entities) contribute most to the current health anomaly from massive amounts of multi-source data. Tracing the root cause of the anomaly refers to inferring the most fundamental triggering reason by following the relational edges defined in the network state semantic graph, starting from the located anomalous features.

[0105] For example, when the health score drops sharply, the anomaly detection submodule can calculate the attention weights or gradient contributions of each neuron in the model (corresponding to different data features), thereby locating the key feature causing the score drop, such as "a surge in core area network latency". For instance, in the semantic graph, starting with "core area network latency", a reverse traversal can be performed along the relationship edges such as "caused by" and "affected by", which may reveal that the root cause is "a hardware failure of a backbone switch", thus generating a tracing chain of "switch failure  network latency  video stuttering  health score decline".

[0106] Optionally, the anomaly detection submodule can be a VAE-based anomaly detection submodule, realizing health assessment, anomaly feature source localization, and root cause tracing. The variational autoencoder (VAE) can be a generative model, used in this application for anomaly detection, which identifies anomaly feature sources by reconstructing input data and calculating reconstruction error.

[0107] In this embodiment, by combining the interpretability analysis capability of the health assessment model with the logical reasoning capability of the external semantic graph, a leap from "perceiving anomalies" to "diagnosing root causes" is achieved. It can not only give the conclusion that the network is unhealthy, but also more accurately point out "where is unhealthy" and "why is it unhealthy", which greatly improves the efficiency of operation and maintenance and the depth of understanding of network status, and enhances the comprehensive perception capability.

[0108] In one exemplary embodiment, the method further includes:

[0109] When the health status representation network has a pre-defined mild abnormality, an optimized intervention action is selected based on a pre-defined reinforcement learning algorithm.

[0110] Simulate and optimize intervention actions in a digital twin environment, and quantify the changes in health status before and after the intervention using a health assessment model;

[0111] When the results of changes in health status indicate that the video surveillance network is in good health, the optimized intervention actions are applied to the video surveillance network to improve its health.

[0112] The reinforcement learning algorithm refers to a machine learning method that learns optimal decision-making strategies by having an agent interact with its environment and learns based on the rewards received. The optimized intervention action refers to measures taken to improve network health, such as "restarting the service," "adjusting routes," or "switching to a backup link." In this application, it is used for dynamic weight adjustment and intervention strategy optimization. The digital twin environment refers to a virtual simulation environment created by high-fidelity modeling of a physical video surveillance network, capable of mapping the state and behavior of physical entities in real time. In this application, it is used to simulate and verify the effectiveness of the intervention strategy.

[0113] For example, when the server detects a minor anomaly (such as a health score within a warning range), the reinforcement learning agent can propose one or more candidate intervention schemes from the action space based on the current network state (e.g., "switch the video stream from camera A from path 1 to path 2"). This optimized intervention is executed in the digital twin environment, and a pre-trained health assessment model is used to predict the change in network health after the action. If the simulation results show that the health can recover to the normal range, this intervention can be distributed to the physical network for execution, achieving closed-loop optimization.

[0114] In this embodiment, a closed-loop autonomous system of "perception-decision-simulation-execution" is constructed by introducing reinforcement learning and digital twin technologies. This allows for the safe and low-cost verification of the effectiveness of optimization strategies in a virtual space, avoiding the risk of secondary failures that may arise from trial and error directly in the physical network. This not only enhances the ability to perceive network status but also extends to intelligent maintenance capabilities based on precise perception, thereby achieving proactive protection and continuous improvement of network health.

[0115] In some exemplary embodiments, such as Figure 5 As shown, Figure 5 A flowchart illustrating reinforcement learning-driven active intervention strategy optimization, provided for embodiments of this application, includes four core modules:

[0116] The State Awareness Module: "Network State Input", which includes three sub-items: "Health Score", "Abnormal Feature Source (such as Link Congestion)" and "Scenario Information", and outputs "Current State S".

[0117] Action selection module: Policy network, receives state S, outputs "optional intervention action A (such as adjusting routing / increasing bandwidth / optimizing coding)".

[0118] Environment interaction module: Network environment, receives action A and executes it, outputs "post-intervention status S'" and "reward R (e.g., health score increased by 20 points / cost -5)".

[0119] Policy update module: The rectangle below is labeled "Reinforcement Learning Update". It receives S, A, R, and S' and feeds them back to the policy network.

[0120] In one exemplary embodiment, the method further includes:

[0121] Based on entity relationships in network state semantic graphs and multi-source video surveillance data, scene types are identified;

[0122] Adjust the model parameters of the health assessment model and the health threshold used for health status assessment according to the scenario type to improve the universality of video surveillance networks of different scales and environments.

[0123] The "scenario type" refers to the specific working environment or business model of the video surveillance network, such as "traffic intersection monitoring," "park security," and "large-scale event support." The normal behavior patterns and health standards of the network may differ significantly across different scenarios. "Model parameters" refers to the adjustable internal weights in the health assessment model. The "health threshold" refers to the critical value used to classify health, sub-health, and abnormal states.

[0124] For example, by analyzing the density and type distribution of devices, network traffic patterns, and user operation characteristics in the semantic graph, the server can automatically identify whether the current network is used for "low-light nighttime monitoring" or "high-mobility area monitoring." For the "nighttime monitoring" scenario, the attention given to quality indicators such as image noise and infrared light status can be increased (adjusting the weights of corresponding features within the model), and the threshold limit for peak network traffic can be appropriately relaxed, since nighttime traffic is inherently lower.

[0125] In this embodiment, the application possesses strong generalization capabilities through a scenario-adaptive parameter and threshold adjustment mechanism. It can understand the differences in network operating states under different application scenarios and dynamically adjust its evaluation criteria, thereby avoiding a "one-size-fits-all" evaluation strategy. This ensures comprehensive and accurate perception and evaluation of network operating states in specific scenarios, whether in a small campus or a large city-level monitoring network.

[0126] In some exemplary implementations, the dynamic weight adaptive fusion algorithm is based on reinforcement learning and automatically adjusts the weights of multi-source data features for different scenarios, thereby enhancing the adaptability of the fused data to complex scenarios.

[0127] For example, in a city's traffic monitoring network, a knowledge graph is constructed, treating traffic intersection cameras and aggregation switches as entities, and establishing relationships such as "connection" and "cause." During data fusion, if a switch has an unusually strong correlation with multiple camera videos, its operational status data weight is increased. For peak and off-peak scenarios, reinforcement learning algorithms automatically adjust data weights, with network traffic data weighting at 60% during peak hours and decreasing to 40% during off-peak hours.

[0128] like Figure 6 As shown, Figure 6 This application provides a schematic diagram of a scenario-based dynamic adaptation architecture, wherein:

[0129] Central Adaptation Module: Dynamic parameter adjustment engine that adaptively adjusts parameters according to scene type.

[0130] Left branch (adapted to size):

[0131] The upper branch is a small network (such as a community), which includes two sub-items: "Data fusion simplification (reducing log data)" and "Lightweight model (reducing the number of GNN layers by 50%)".

[0132] The lower-level branch is a large-scale network (such as a city), which includes two sub-items: "full data fusion" and "enhanced model (increasing the number of LSTM units)".

[0133] Right-hand branch (adapted to different environments):

[0134] The upper branch is for indoor environments (such as parks), which includes two sub-items: "Standard anti-interference processing" and "Conventional health threshold (packet loss rate ≤5%)".

[0135] The lower branch is "Field environment (such as oil field)", which includes two sub-items: "Adaptive filtering anti-interference" and "Relaxed threshold (packet loss rate ≤10%)".

[0136] All branches are aggregated into the "Post-Adaptation Execution Module," which outputs the post-adaptation assessment and intervention strategies.

[0137] In some exemplary embodiments, this application provides the following fusion structure, including:

[0138] Reinforcement learning: a branch of machine learning that learns optimal policies through interaction with the environment in a "state-action-reward" pattern. This invention is used for dynamic weight adjustment and intervention policy optimization.

[0139] Digital twin: A digital mirror image of a physical entity that can map the state and behavior of the physical entity in real time. In this invention, it is used to simulate and verify the effectiveness of intervention strategies.

[0140] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0141] The following describes the health status determination device for a video surveillance network provided in the embodiments of this application. The health status determination device for a video surveillance network has the same inventive concept as the above-described health status determination method for a video surveillance network. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the health status determination device for a video surveillance network provided below can be referred to the limitations of the health status determination method for a video surveillance network described above. The health status determination device for a video surveillance network described below and the health status determination method for a video surveillance network described above can be referred to each other, and will not be repeated here.

[0142] In one exemplary embodiment, Figure 7 This is a schematic diagram of the structure of a health status determination device for a video surveillance network provided in an embodiment of this application, as shown below. Figure 7 As shown, the health status determination device 70 of the video surveillance network includes: ..., ..., ..., ..., where:

[0143] The data acquisition module 710 is used to acquire multi-source video surveillance data from the video surveillance network. The multi-source video surveillance data includes equipment and operating status data, network traffic data, video quality data, and user operation log data.

[0144] The semantic fusion module 720 is used to perform semantic fusion on multi-source video surveillance data based on a pre-built network state semantic graph to obtain semantic association features;

[0145] The health assessment module 730 is used to assess the health status of the video surveillance network based on semantic association features and a pre-trained health assessment model. The health assessment model is trained using a network architecture that integrates graph neural networks and long short-term memory networks.

[0146] The anomaly tracing module 740 is used to determine the anomaly tracing results through a health assessment model when the health status characterizes an anomaly in the video surveillance network.

[0147] In an exemplary embodiment, the data acquisition module 710 is further configured to construct a semantic graph for the video surveillance network, using devices, links, fault types and video quality indicators as entities, and establishing entity relationships between entities through relational edges; and constructing a network state semantic graph based on each entity and entity relationships.

[0148] In an exemplary embodiment, the data acquisition module 710 is used to determine the weight parameters corresponding to the multi-source video surveillance data; and to perform weighted fusion of the multi-source video surveillance data according to the entity relationships of the network state semantic graph and the weight parameters to obtain semantic association features.

[0149] In an exemplary embodiment, the health assessment module 730 is used to input semantic association features into the health assessment model, learn the spatial dependency relationship between entities through a graph neural network, and capture the temporal change features of the network state through a long short-term memory network; it integrates the spatial dependency relationship and the temporal change features to output a health score, which represents the health status.

[0150] In an exemplary embodiment, the anomaly tracing module 740 is used to locate abnormal data features based on the health score through the anomaly detection submodule built into the health assessment model; and to trace the root cause of the anomaly based on the abnormal data features and the network state semantic graph to obtain the anomaly tracing result.

[0151] In an exemplary embodiment, the health assessment module 730 is used to select an optimized intervention action based on a preset reinforcement learning algorithm when the health status representation network has a preset mild abnormality; simulate the optimized intervention action in a digital twin environment, quantify the health status change results before and after the intervention through the health assessment model; and apply the optimized intervention action to the video surveillance network when the health status change results indicate that the video surveillance network is healthy and normal, so as to improve the health of the video surveillance network.

[0152] In an exemplary embodiment, the health assessment module 730 is used to identify scene types based on entity relationships in the network state semantic graph and multi-source video surveillance data; and adjust the model parameters of the health assessment model and the health threshold used for health status assessment according to the scene type, so as to improve the universality of video surveillance networks of different scales and environments.

[0153] In one exemplary embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the health status determination method for any of the video surveillance networks described above.

[0154] In one exemplary embodiment, this application also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the health status determination method for any of the video surveillance networks described in the above embodiments.

[0155] In one exemplary embodiment, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the health status determination method for any of the video surveillance networks described in the above embodiments.

[0156] Indicatively, such as Figure 8 As shown, Figure 8 This is a schematic diagram of the internal structure of a computer device 800 provided in an embodiment of this application. The computer device 800 can be provided as a server. (Refer to...) Figure 8 The computer device 800 includes a processing component 802, which further includes one or more processors, and memory resources represented by memory 801 for storing instructions, such as application programs, that can be executed by the processing component 802. The application programs stored in memory 801 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 802 is configured to execute instructions to perform the health status determination method for a video surveillance network according to any of the above embodiments.

[0157] The computer device 800 may also include a power supply component 803 configured to perform power management of the computer device 800, a wired or wireless network interface 804 configured to connect the computer device 800 to a network, and an input / output (I / O) interface 805. The computer device 800 may operate on an operating system stored in memory 801, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0158] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0159] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0160] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0161] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0162] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining the health status of a video surveillance network, characterized in that, The method includes: Acquire multi-source video surveillance data from the video surveillance network, including device and operating status data, network traffic data, video quality data, and user operation log data; Based on a pre-constructed network state semantic graph, the multi-source video surveillance data is semantically fused to obtain semantic association features; Based on the semantic association features, a pre-trained health assessment model is used to assess the health status of the video surveillance network, thereby obtaining the health status of the network; wherein, the health assessment model is trained using a network architecture that integrates graph neural networks and long short-term memory networks. When the health status indicates an abnormality in the video surveillance network, the abnormality source is determined using the health assessment model.

2. The method according to claim 1, characterized in that, Before performing semantic fusion of the multi-source video surveillance data based on a pre-constructed network state semantic graph to obtain semantic association features, the process includes: Construct a semantic graph for video surveillance networks, using devices, links, fault types, and video quality indicators as entities, and establish entity relationships between entities through relation edges; Based on each entity and the relationships between them, a network state semantic graph is constructed.

3. The method according to claim 2, characterized in that, The multi-source video surveillance data is semantically fused based on a pre-constructed network state semantic graph to obtain semantic association features, including: Determine the weight parameters corresponding to the multi-source video surveillance data; Based on the entity relationships in the network state semantic graph and the weight parameters, the multi-source video surveillance data is weighted and fused to obtain the semantic association features.

4. The method according to claim 1, characterized in that, The step of assessing the health status of the video surveillance network based on the semantic association features using a pre-trained health assessment model to obtain the health status includes: The semantic association features are input into the health assessment model, the spatial dependencies between entities are learned through graph neural networks, and the temporal change features of the network state are captured through long short-term memory networks. By integrating the spatial dependency relationship and the temporal change characteristics, a health score is output, which represents the health status.

5. The method according to claim 4, characterized in that, The process of determining the source of anomalies using the health assessment model includes: The anomaly detection submodule built into the health assessment model is used to locate abnormal data features based on the health score. Based on the abnormal data characteristics and the network state semantic graph, the root cause of the anomaly is traced to obtain the anomaly tracing result.

6. The method according to claim 1, characterized in that, The method further includes: When the health status representation network has a preset mild abnormality, an optimized intervention action is selected based on a preset reinforcement learning algorithm. The optimized intervention actions are simulated in a digital twin environment, and the changes in health status before and after the intervention are quantified using the health assessment model. If the changes in health status indicate that the video surveillance network is in good health, the optimized intervention actions are applied to the video surveillance network to improve its health.

7. The method according to claim 1, characterized in that, The method further includes: Based on the entity relationships in the network state semantic graph and the multi-source video surveillance data, the scene type is identified; Based on the scenario type, the model parameters of the health assessment model and the health threshold used for health status assessment are adjusted to improve the universality of video surveillance networks of different scales and environments.

8. A device for determining the health status of a video surveillance network, characterized in that, The device includes: The data acquisition module is used to acquire multi-source video surveillance data from the video surveillance network. The multi-source video surveillance data includes equipment and operating status data, network traffic data, video quality data, and user operation log data. The semantic fusion module is used to perform semantic fusion on the multi-source video surveillance data based on a pre-constructed network state semantic graph to obtain semantic association features; The health assessment module is used to assess the health status of the video surveillance network based on the semantic association features and using a pre-trained health assessment model; wherein the health assessment model is trained using a network architecture that integrates graph neural networks and long short-term memory networks. The anomaly tracing module is used to determine the anomaly tracing result through the health assessment model when the health status indicates an anomaly in the video surveillance network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images