Method for dynamic evaluation of communication quality based on spatio-temporal graph neural network

Abstract

The application relates to the technical field of communication quality evaluation, and discloses a communication quality dynamic evaluation method based on a space-time graph neural network, which comprises the following steps: step 1, collecting the physical coordinates, signal strength, link delay and packet loss rate of a wireless network node in real time; step 2, constructing a space-time graph structure based on an improved Delaunay triangulation algorithm, wherein the physical coordinates of the device are taken as vertices V, and the communication link is taken as an edge E; the edge weight is dynamically calculated; step 3, constructing a three-layer space-time graph convolution network model and generating a quality evaluation matrix Q; and step 4, dividing the quality state interval based on a Gaussian mixture model, and determining the communication quality state of the node based on the quality state interval to which the Q value belongs. The application has strong adaptability and can realize continuous, dynamic and robust evaluation of the communication quality state of the node in a complex communication scene.

Description

Technical Field

[0001] This invention relates to the field of communication quality assessment technology, and specifically to a dynamic communication quality assessment method based on a spatiotemporal graph neural network. Background Technology

[0002] In dynamic and equipment-intensive scenarios such as film and television production and large-scale performances, wireless communication quality assessment is an essential and challenging aspect. In these scenarios, mobile nodes such as cameras, drones, and lighting controllers continuously interact in complex and ever-changing electromagnetic environments. Their communication quality directly affects the execution effectiveness and economic benefits of critical tasks. Related industry reports show that performance interruptions due to communication failures at large-scale events result in annual losses exceeding ten million yuan; while in film and television shooting, the scrap rate of footage caused by wireless image transmission lag can reach as high as 12%.

[0003] However, traditional evaluation methods based on fixed probes or statistical models perform poorly in this type of scenario: on the one hand, rapid device displacement causes network topology to change dozens of times per second, making static modeling methods unable to effectively capture the dynamic characteristics of instantaneous link quality; on the other hand, multi-source interference (including audience mobile hotspots, stage electronic equipment, external radio frequency sources, etc.) is highly coupled in the time and frequency domains, making it difficult for existing technologies to effectively separate the specific impacts of various interference sources on communication quality. Existing technologies lack the ability to jointly model multi-dimensional dynamic factors such as spatial location, spectrum status, and device movement, forcing field engineers to rely solely on experience for parameter tuning—not only inefficient but also unable to prevent sudden communication collapses.

[0004] Current research explores improvement along two main paths, but both have significant limitations. The first path is a spectrum sensing scheme driven by cognitive radio technology, which essentially avoids interference by scanning environmental frequency bands in real time and combining historical databases. This scheme works well for scenarios with known interference sources, but it suffers from significant response delays to unknown sudden interference (such as temporarily activated non-standard equipment) and completely fails to consider the sudden signal attenuation caused by equipment movement. The second path is the application of graph neural networks in the field of communication, which involves abstracting the network into nodes and edges for quality prediction. However, current technologies only support static topology—for real-time equipment displacement common in film and television shooting (such as camera track movement and changes in drone aerial shooting paths), manual remodeling is still required. A more critical drawback is that existing methods generally neglect spatial scale characteristics: the dense collision of equipment in the core area of ​​the stage and the signal attenuation in the distant audience area require differentiated modeling, while current single-scale graph convolution cannot capture instantaneous interference or track trend evolution, resulting in a high false negative rate. Summary of the Invention

[0005] The present invention aims to provide a dynamic evaluation method for communication quality based on spatiotemporal graph neural networks, which solves the technical problems of low efficiency and inability to effectively handle dynamic evaluation of complex channels in existing evaluation techniques. It is highly adaptable and can achieve continuous, dynamic and robust evaluation of the communication quality status of nodes.

[0006] The basic solution provided by this invention is: a dynamic evaluation method for communication quality based on spatiotemporal graph neural networks, comprising the following steps:

[0007] Step 1: Real-time collection of physical coordinates, signal strength, link latency, and packet loss rate of wireless network nodes;

[0008] Step 2: Construct the spatiotemporal graph structure based on the improved Delaunay triangulation algorithm Where: the physical coordinates of the device are vertices V and the communication link is an edge E;

[0009] Dynamically calculate edge weights ;in, and These are the weighting coefficients. , ; This is the baseline link delay value. The baseline packet loss rate;

[0010] Step 3: Construct a three-layer spatiotemporal graph convolutional network model and generate the quality evaluation matrix Q;

[0011] Step 4: Divide the quality state intervals based on the Gaussian mixture model, and determine the communication quality state of the node based on the quality state interval to which the Q value belongs.

[0012] The working principle and advantages of this invention are as follows:

[0013] This invention presents a dynamic communication quality assessment method based on spatiotemporal graph neural networks. This method is highly adaptable and can achieve continuous, dynamic, and robust assessment of node communication quality status in complex communication scenarios. The key points are:

[0014] First, this solution can incorporate complex spatiotemporal dynamic characteristics into a unified analysis framework, enabling dynamic evaluation of node communication quality.

[0015] Specifically, this scheme establishes a precise mapping relationship between physical space and communication state, completely breaking through the limitations of traditional static modeling. First, it introduces physical coordinates, creating conditions for accurately depicting the spatial topological relationships between nodes. This ensures that the evaluation model can perceive the impact of changes in the network's physical layout, such as signal quality fluctuations caused by node movement or the appearance of obstacles, providing solid support for dynamic evaluation. Then, dynamic parameters such as device location, link quality, and motion trajectory are transformed into vertex and edge attributes of a spatiotemporal graph structure, enabling multi-scale features to be jointly modeled. Combined with an improved Delaunay triangulation to construct the spatiotemporal graph structure, it cleverly transforms the network physical topology into a mathematical graph model, ensuring the relative uniformity and stability of the graph structure. Through this mechanism of dynamically constructing the graph and calculating weights based on physical coordinates, the model can respond in real time to dynamic factors such as node movement and link state changes. The generated graph structure has spatiotemporal evolution capabilities, helping to capture continuous changes in network state and achieve dynamic analysis.

[0016] Second, this scheme constructs a decoupled extraction mechanism for multi-scale spatiotemporal features, which helps to improve the accuracy of communication quality characterization.

[0017] Specifically, existing graph neural networks typically use a single convolutional kernel to process global data, while this solution innovatively designs a hierarchical convolutional architecture that can specifically process data at different scales, achieving differentiated modeling and helping to capture transient interference and track trend evolution. Furthermore, the spatiotemporal graph convolutional network has the ability to simultaneously capture spatial dependencies and temporal correlations. At the graph structure level, graph convolutional operations can effectively aggregate information from neighboring nodes, learning complex spatial association patterns between nodes due to physical distance, signal interference, routing paths, etc. In the temporal dimension, historical sequences of key indicators such as link latency and packet loss rate can be analyzed through sequence learning mechanisms to identify temporal evolution patterns such as network traffic congestion and periodic fluctuations in channel quality. Based on the joint extraction and fusion of spatiotemporal features, the model can understand the complete trajectory of network state changes, such as the gradual process of signal attenuation caused by node movement, or the instantaneous congestion diffusion phenomenon caused by sudden traffic surges. The resulting quality assessment matrix Q contains rich spatiotemporal context information, enabling a more comprehensive representation of communication quality.

[0018] Third, this scheme utilizes a Gaussian mixture model to achieve a continuous, dynamic, and robust assessment of the communication quality status of nodes, which helps to improve the accuracy of communication quality assessment.

[0019] Specifically, this solution uses a Gaussian Mixture Model (GMM) to divide the quality assessment matrix Q into state intervals, which conforms to the actual network state distribution characteristics. Communication quality states (such as excellent, good, average, and poor) are usually continuously changing and have blurred boundaries, which cannot be clearly separated by a simple threshold. This solution uses a Gaussian Mixture Model, which, through a linear combination of multiple Gaussian distributions, can flexibly fit the multimodal distribution characteristics that the quality assessment values ​​may exhibit, and automatically discover potential natural clusters in the data. In this solution, the division of state intervals is not based on a subjectively preset fixed threshold, but is driven by the statistical characteristics of the data itself, which better reflects the inherent structure of quality states in real scenarios. This makes the judgment of node communication quality more robust, can tolerate normal fluctuations in assessment values, and reduces fluctuations caused by small changes leading to frequent jumps in state categories. Its output results are more stable and more consistent with the understanding of state judgment in actual operation and maintenance. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of an embodiment of the communication quality dynamic evaluation method based on spatiotemporal graph neural network of the present invention. Detailed Implementation

[0021] The following detailed explanation illustrates the specific implementation methods:

[0022] The basic implementation examples are as follows: Figure 1 As shown: A dynamic evaluation method for communication quality based on spatiotemporal graph neural networks includes the following steps:

[0023] Step 1: Real-time acquisition of the physical coordinates and signal strength of wireless network nodes. Link delay and packet loss rate The physical coordinates refer to three-dimensional physical coordinates. .

[0024] The wireless network nodes include communication nodes, sensing nodes, and dynamic nodes; the communication nodes include master control devices, terminal execution devices, and relay forwarding devices; the sensing nodes include spectrum monitoring devices and interference source marking devices; and the dynamic nodes include mobile carrier devices and temporary access devices.

[0025] Specifically, in this embodiment, the main control equipment includes a concert control console, a video broadcasting vehicle, etc.; the terminal execution equipment includes cameras, drones, lighting controllers, etc. The relay forwarding equipment includes stage edge repeaters, filming site mesh nodes, etc.

[0026] The spectrum monitoring equipment includes a distributed spectrum sensor; the interference source marking equipment includes a mobile phone hotspot detector in the audience area. The mobile carrier equipment includes a camera crane, a drone-borne relay, etc.; the temporary access equipment includes a staff handheld tablet, a temporary lighting controller, etc.

[0027] This solution is applicable to channel dynamic evaluation scenarios involving multiple types of complex wireless devices, such as film and television shooting and large-scale performances.

[0028] Step 2: Construct the spatiotemporal graph structure based on the improved Delaunay triangulation algorithm Where: the physical coordinates of the device are vertices V and the communication link is an edge E.

[0029] Specifically, vertex set The device is represented by coordinates. Sure.

[0030] The edge set E is generated using an improved Delaunay triangulation algorithm, satisfying the empty circumcircle criterion.

[0031] The improved Delaunay triangulation algorithm includes:

[0032] Based on the original Delaunay triangulation algorithm, a link quality constraint is added—topology connections are established only when the signal-to-noise ratio between wireless network nodes is >10dB. A time decay factor is also introduced, causing edge weights to decay according to a set time decay factor as the link idle time increases; that is, if the edge weight increases... exist If no data is transmitted within a minute, its weight will be adjusted accordingly. Updated to . This is the time decay factor.

[0033] By using a signal-to-noise ratio threshold constraint, invalid connections can be effectively eliminated, retaining only actually communicable links. This ensures the graph structure accurately reflects the true channel connectivity and resolves the problem of physically adjacent but unreachable links (such as signal shielding caused by the metal frame of a concert stage). Introducing an exponential decay function to update the weights of idle links avoids topology rigidity caused by temporary link outages (such as brief offline status of mobile devices during filming). The set time decay factor ensures that when there is no data for 30 minutes, the corresponding weight drops to below 1% of the initial value, and when there is no data for 5 minutes, the corresponding weight remains at 37%, balancing response speed and stability.

[0034] Dynamically calculate edge weights ;Right now: .

[0035] in, and These are the weighting coefficients. , ; This is the baseline link delay value. The baseline packet loss rate.

[0036] And, when hour, Take 1.2 Based on this setting, in scenarios that are extremely sensitive to latency (such as wireless image transmission for film and television, and real-time control command transmission for performances), once the latency exceeds the threshold, the impact of latency on link quality assessment can be highlighted in a timely manner, prioritizing the protection of low latency requirements and reducing the adverse impact of high latency on services.

[0037] Step 3: Construct a three-layer spatiotemporal graph convolutional network model and generate a quality evaluation matrix Q.

[0038] Includes the following sub-steps:

[0039] S3.1, Construct small-scale layers; subgraphs with spatiotemporal graph structures. Using spatial graph convolution and temporal convolution as input, perform convolution on the spatial graph. . It represents a node in the original spatiotemporal graph structure (such as a network node corresponding to a camera, drone, or other device in a wireless communication scenario). The central node (corresponding to the selected main camera node and the stage center device node) is used as a reference point for distance judgment.

[0040] By using a small-scale layer, transient interference between devices (such as collisions in camera group control signals) can be effectively captured.

[0041] S3.2, Construct a mesoscale layer; using a subgraph with a spatiotemporal graph structure. Using the input as input, perform spatiotemporal separation convolution and calculate the decay gradient. Wherein, .

[0042] The nonlinear attenuation trend of the signal with distance can be characterized by the mesoscale layer.

[0043] S3.3, Construct large-scale layers; a full graph with a spatiotemporal graph structure. As input, a multi-head attention mechanism is executed and the input graph readout layer is used.

[0044] Through large-scale layers, cross-regional latent interference can be identified (such as frequency band coupling between lighting equipment 500 meters away and the main camera).

[0045] S3.4 aligns and concatenates the output features from S3.1 to S3.3, and then performs spatiotemporal feature fusion through a gated recurrent unit to generate the quality evaluation matrix Q. .

[0046] When performing feature stitching, the output features of S4.1 to S4.3 are also stitched together with historical data; the historical data includes the average interference intensity per hour over the past 24 hours.

[0047] This innovative solution employs a hierarchical convolutional architecture: small-scale layers focus on capturing instantaneous collisions in densely populated areas, medium-scale layers analyze signal attenuation patterns using gradient operators, and large-scale layers utilize multi-head attention to uncover implicit cross-regional correlations. Through the strong coupling between physical scale and algorithmic structure, this solution effectively resolves the inherent contradiction in traditional methods where microsecond-level impulse noise and minute-level interference trends are incompatible.

[0048] Step 4: Divide the quality state intervals based on the Gaussian mixture model, and determine the communication quality state of the node based on the quality state interval to which the Q value belongs.

[0049] Specifically, the quality state space is obtained based on the clustering of historical Q values, including: normal state interval, Q>0.8; warning state interval, 0.6≤Q≤0.8; and fault state interval, Q<0.6.

[0050] This step also includes an early warning mechanism and a fault response mechanism.

[0051] The early warning processing mechanism includes: when the node Q value falls into the early warning state range within 3 consecutive sampling periods, the LSTM frequency band prediction module is activated to predict the availability probability of each frequency point in the future time domain.

[0052] The inputs to the LSTM band prediction module include: current spectrum sensing data, device location change vector, historical quality assessment matrix, and interference source characteristics.

[0053] The fault response mechanism includes: when the packet loss rate is >30%, switching the wireless network node to the optimal frequency predicted by the LSTM band prediction module; when Q <0.6 and lasts for 10 seconds, generating a spectrum heatmap and node association strength map in the HSV color space.

[0054] The steps for generating the spectral heatmap include: mapping the Q matrix to the HSV color space, wherein: hue H is positively correlated with the center frequency of the frequency band; saturation S is positively correlated with signal strength; and brightness V is positively correlated with Q value.

[0055] The node association strength graph uses edge transparency Characterizes communication quality. ;in, and These are the node feature vectors.

[0056] In this step, unlike fixed threshold early warning systems, this solution generates three levels of thresholds that adapt to the network state based on a Gaussian mixture model, and combines LSTM frequency availability prediction to achieve hierarchical response: single-frequency interference executes millisecond-level seamless switching, and full-band collapse triggers a visualized degradation scheme, which can transform the traditional passive response into active defense and significantly improve the success rate of frequency band migration.

[0057] In addition, in this step, the frequency migration results are fed back to the historical database; and incremental learning is used to update the LSTM model weights of the LSTM band prediction module.

[0058] Existing research often separates spectrum sensing from network modeling. This solution, however, drives dynamic updates of edge weights through real-time spectrum data, uses spatiotemporal graph features to guide frequency selection strategies, and feeds the transfer results back to the LSTM model to form an incremental learning loop. This helps solve the problem of the disconnect between perception and decision-making in existing technologies. Taking a concert scenario as an example, when audience members' mobile phone hotspots cause local frequency band congestion, this solution can not only locate affected camera nodes through graph convolution but also automatically switch to the optimal frequency based on spectrum prediction. Simultaneously, it records interference patterns to enhance responsiveness in similar future scenarios, helping to reduce maintenance downtime.

[0059] Step 5: If there are node changes in the wireless network nodes, re-collect node data and update the spatiotemporal graph structure to generate a new quality assessment matrix in order to re-determine the communication quality status of the nodes.

[0060] The method for determining node changes includes: real-time monitoring of the actual movement distance of each node based on its physical coordinates in the wireless network; when the actual movement distance... When the value exceeds a preset threshold, topology reconstruction is triggered.

[0061] In this embodiment, or At 0.15, topology reconstruction is triggered. Topology reconstruction refers to rebuilding the spatiotemporal graph structure (i.e., re-collecting node data and updating the spatiotemporal graph structure) to maintain the consistency of the graph structure with the real wireless environment.

[0062] In topology reconstruction, if only local wireless network nodes change, only the subgraph within a 10m radius of the displaced node is reconstructed; if more than 50% of the wireless network nodes change, Delaunay triangulation is performed on all nodes, and then the edge weights are updated.

[0063] This step combines a displacement-triggered mechanism to achieve adaptive topology updates, which solves the resource waste problem of fixed-period reconstruction. Furthermore, in scenarios where film and television shooting equipment moves frequently, topology reconstruction takes less time. This dynamic modeling capability also helps to further improve the accuracy of communication quality assessment.

[0064] This embodiment provides a dynamic communication quality assessment method based on spatiotemporal graph neural networks. It dynamically constructs a spatiotemporal graph and calculates adaptive link weights through improved Delaunay triangulation. Combined with a spatiotemporal graph convolutional network, it simultaneously captures the spatial topological dependencies and temporal evolution of the communication network. Finally, it uses a Gaussian mixture model to achieve a continuous, dynamic, and robust assessment of the communication quality status of nodes, which helps to improve the efficiency and accuracy of communication quality assessment.

[0065] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for dynamic evaluation of communication quality based on a spatio-temporal graph neural network, characterized in that, The method comprises the following steps: Step 1, collecting the physical coordinates, signal strength, link delay and packet loss rate of the wireless network nodes in real time; Step 2: Constructing the spatio-temporal graph structure based on the improved Delaunay triangulation algorithm wherein: the device physical coordinates are vertices V, and the communication links are edges E; Dynamically calculating edge weights ; wherein, and are weight coefficients, , ; is a reference link delay value, is a reference packet loss rate; The improved Delaunay triangulation algorithm comprises: Based on the original Delaunay triangulation algorithm, a link quality constraint condition is added, that is, a topological connection is established only when the signal-to-noise ratio between the wireless network nodes is greater than 10 dB; and a time decay factor is introduced, so that the edge weight decays according to a set time decay factor as the link idle time increases; Step 3, constructing a three-layer space-time graph convolution network model and generating a quality evaluation matrix Q; The step 3 comprises the following sub-steps: S3.1, construct small-scale layer; subgraph of spatiotemporal graph structure for input, perform spatial graph convolution and temporal convolution; for vertex set of small-scale layer; for edge set of small-scale layer; ; S3.2, construct meso-scale layers; subgraph of spatiotemporal graph structure For input, perform spatiotemporal separable convolution and compute decay gradient; For vertex set of meso-scale layers; For edge set of meso-scale layers; ; representing a node in the original spatiotemporal graph structure; for the central node, a reference point for distance judgment; S3.3, construct large-scale layers; full graph of spatiotemporal graph structure For input, perform multi-head attention mechanism and input graph reading layer; S3.4, aligning and splicing the output features of S3.1-S3.3, and generating a quality evaluation matrix Q through a gating cycle unit for space-time feature fusion; Step 4, dividing the quality state interval based on the Gaussian mixture model, and determining the communication quality state of the node based on the quality state interval to which the Q value belongs.

2. The method of claim 1, wherein, The wireless network nodes comprise communication nodes, sensing nodes and dynamic nodes; the communication nodes comprise master control devices, terminal execution devices and relay forwarding devices; the sensing nodes comprise spectrum monitoring devices and interference source marking devices; and the dynamic nodes comprise mobile carrier devices and temporary access devices.

3. The method of claim 1, wherein, In S3.4, when the feature splicing is performed, the output features of S3.1-S3.3 are also spliced with historical data; the historical data include the average interference intensity of each hour in the past 24 hours.

4. The method of claim 1, wherein, The quality state interval is obtained based on clustering of historical Q values, and comprises: a normal state interval Q>0.8; a warning state interval 0.6≤Q≤0.8; and a fault state interval Q<0.

6.

5. The method of claim 4, wherein, In step 4, a warning processing mechanism is further provided: When the Q value of the node falls into the warning state interval for three consecutive sampling periods, an LSTM frequency band prediction module is started to predict the availability probability of each frequency point in the future time domain; The input of the LSTM frequency band prediction module comprises: current spectrum sensing data, device position change vector, historical quality evaluation matrix and interference source feature.

6. The method of claim 5, wherein, In step 4, a fault response mechanism is further provided: when the packet loss rate is greater than 30%, the wireless network node is switched to the optimal frequency point predicted by the LSTM frequency band prediction module; and when Q<0.6 and lasts for 10 seconds, a spectrum heat map in HSV color space and a node association intensity map are generated.

7. The method of claim 6, wherein, The generation steps of the spectrum heat map comprise: mapping the Q matrix to the HSV color space, wherein: the hue H is positively correlated with the center frequency of the frequency band; the saturation S is positively correlated with the signal strength; and the lightness V is positively correlated with the Q value.

8. The method of claim 1, wherein, Further comprising step 5, if there is a node change in the wireless network node, the node data is re-collected and the space-time graph structure is updated to generate a new quality evaluation matrix to re-determine the communication quality state of the node.

Images