A communication base station traffic analysis method based on graph embedding technology and dynamics equation reasoning

Abstract

The application discloses a communication base station traffic analysis method based on graph embedding technology and dynamics equation reasoning, which comprises the following steps: measuring traffic data information of a communication base station network; using graph embedding technology to reduce the dimension of a high-dimensional sparse adjacency matrix A eigenvector space, and projecting time-varying traffic data to a low-dimensional characteristic space; constructing two basic function libraries; reasoning network dynamics equations from the basic function library based on the reduced data space. The application uses the time-varying property of the dynamics equation to capture the self-evolution and interaction effect of the network node, and reasons the equation from the basic function library, so that the inherent characteristics and potential rules of the communication base station traffic can be fully captured, the explainability is relatively high, and better precision can be achieved for short-term and long-term traffic prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence prediction algorithms, and in particular to a communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning. Background Art

[0002] With the rapid development of wireless communication technology and the widespread adoption of mobile devices, explosive growth in network traffic has become inevitable. Communication base stations serve as information transfer hubs between mobile devices and are also the hubs along the network traffic transmission path. Allocating network and human resources to these base stations presents a significant challenge for network operators. The widespread implementation of 5G communication technology will spur the emergence of even more internet application scenarios, such as autonomous driving, virtual reality, and the metaverse, further diversifying demand for mobile data.

[0003] While changes in communication network traffic intensity are highly dynamic over time, they also reveal underlying trends: the distribution of base station traffic intensity within a region exhibits specific regularities. Deducing the mathematical characteristics underlying these periodic variations in base station traffic is crucial for solving a range of problems, including base station site selection, network resource allocation, and internet user profiling.

[0004] Currently, a large number of experts and scholars have conducted relevant research on the highly dynamic and periodic characteristics of base station traffic. Traditional base station traffic models are mainly aimed at prediction, and representative models include time series models, Poisson models, and Markov models. Li et al. (The Learning and Prediction of Application-level Traffic Data in Cellular Networks, 2017) proposed the α-stable model based on both time and space dimensions. This model significantly improves accuracy compared to traditional linear time series models. Kusdarwati et al. (System for Prediction of Non-Stationary Time Series based on the Wavelet Radial Bases Function Neural Network Model, 2018) proposed the SARI-MA model, which analyzes the autocorrelation of time series and can capture the seasonal characteristics of network traffic. Cao et al. (Network traffic analysis and prediction of hotspot in cellular network, 2018) used a triple exponential smoothing model to predict uplink and downlink traffic data. Le et al. (Applying Big Data, Machine Learning, and SDN / NFV for 5G Early-Stage Traffic Classification and Network QoS Control, 2018) used both linear and nonlinear methods to model and predict base station traffic. The results showed that the nonlinear method outperformed the linear method. Although the parameterized statistical model described above can describe the nonlinear characteristics of base station traffic, it still suffers from problems such as low model fitting accuracy and poor generalization ability.

[0005] In recent years, with the widespread application of artificial intelligence technology in various fields, deep learning, as a typical non-parametric model, has gradually become a hot research direction in traffic prediction. Liu et al. (Short-term traffic flow prediction with Conv-LSTM, 2017) combined convolutional networks and LSTM to extract the spatiotemporal correlations of traffic data. Wang et al. (Bayesian optimization of support vector machine for regression prediction of short-term traffic flow, 2019) used support vector regression to solve the learning problem under small sample traffic data, but the generalization ability for large sample cases is weak. Gao et al. (Short-term Traffic Flow Prediction Based on Time-space Characteristics, 2020) designed a gradient boosting method for traffic prediction and used convolutional neural networks to effectively extract the spatial characteristics of traffic. Yuan et al. (Prediction Method of Wireless Network Traffic Based on Spatiotemporal Features, 2022) introduced an attention mechanism based on convolutional neural networks to improve the extraction of global spatial correlations. It is true that deep learning can effectively extract the spatiotemporal characteristics of traffic data, and the prediction accuracy within a certain time range is satisfactory, but the high complexity of the model will also lead to a decrease in model interpretability.

[0006] Current mainstream methods for base station traffic analysis can provide reasonable and effective predictions of future traffic. However, they are less effective at explaining underlying trends in traffic. The inherent "black box" nature of deep learning prevents models from effectively analyzing and interpreting extracted features. Furthermore, due to the high dimensionality of the adjacency matrix representing node connections in complex networks, most models suffer from a large number of parameters and complex structures, making it extremely difficult to reason about base station traffic trends. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning. The topological laws of the base station network are captured by establishing an adjacency matrix representing the degree of proximity between base stations, and the adjacency matrix is ​​reduced in dimension using graph embedding technology based on Laplace eigenmap. The time-varying traffic data is projected into a low-dimensional feature space, and a basic function library that may generate real network dynamics is constructed to infer the network dynamic equation that best matches the evolution of the system state.

[0008] To solve the above technical problems, the present invention provides a communication base station traffic analysis system based on graph embedding technology and dynamic equation reasoning, comprising: a communication base station information acquisition module, an input agent module, a feature space dimensionality reduction module, a dynamic model inference module and a function library module;

[0009] The communication base station information collection module is composed of the communication base station's locator and traffic information detector to form an information collection network. The locator records the base station's spatial information, and the traffic information detector captures the direction and intensity of time-varying traffic.

[0010] The input agent module encodes the network data input by the communication base station information acquisition module into a matrix form and inputs it into the feature space dimensionality reduction module;

[0011] The feature space dimensionality reduction module fully extracts spatiotemporal features to achieve data dimensionality reduction, and inputs the low-dimensional data into the dynamic model inference module;

[0012] The dynamic model inference module searches for the most explanatory and concise function model in the function library module;

[0013] The library module stores the basic functions that may generate real network dynamics.

[0014] Accordingly, a communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning includes the following steps:

[0015] Step 1: Measure the traffic data of the communication base station network and represent the communication base station network as a weighted acyclic graph D = (V, E, A), where V is the node set consisting of the communication base stations, E is the edge set between the nodes, and A is the adjacency matrix that measures the distance between the nodes.

[0016] The calculation method of the weighted adjacency matrix A, which represents the node proximity, is:

[0017]

[0018] A ij Indicates base station v with number i i With base station v numbered j j The distance weight, ρ(v i , v j ) represents the base station v numbered i i With base station v numbered j j The distance between 2 is the variance of the distance between base stations, is the set distance threshold;

[0019] Step 2: Use graph embedding technology to reduce the dimensionality of the eigenvector space of the high-dimensional and sparse adjacency matrix A and project the time-varying traffic data into a low-dimensional feature space;

[0020] Step 3: Build two basic function libraries;

[0021] Step 4: Infer the network dynamics equations from the basic function library based on the reduced-dimensional data space.

[0022] Preferably, in step 2, using graph embedding technology to reduce the dimensionality of the eigenvector space of the high-dimensional sparse adjacency matrix A, and projecting the time-varying traffic data into the low-dimensional feature space specifically includes the following steps:

[0023] Step 21: Use the mapping function F(·) to transform each eigenvector Convert to low-dimensional representation Make Y i (t) = F(X i (t)), M is the dimension of the original feature space, and d is the dimension of the transformed feature space;

[0024] Step 22: Calculate the Laplace matrix L≡DA. The diagonal matrix D stores the degree of the base station node.

[0025] Step 23: Given the constraint Y T DY=E, find the optimal function F(·) under the constraints:

[0026]

[0027] Preferably, in step 23, given constraint condition Y T DY=E, find the optimal function F(·) under the constraints:

[0028]

[0029] After using linear projection for dimensionality reduction, the process of solving the coefficients includes the following steps:

[0030] Step 231: Use Y=X T The linear projection of w will be down to

[0031] Step 232: Use linear projection Y=X T w , the objective function becomes:

[0032]

[0033] Step 233: Under the constraint condition ω T XT When DXω=E, the Lagrangian function is constructed as:

[0034]

[0035] Among them, Λ is a diagonal matrix, and the elements Λ on the diagonal ii =λ i is the Lagrange multiplier;

[0036] Step 234: Use the generalized eigenvalue method to solve the Lagrange equation to obtain the coefficient ω of the linear projection. The equation is in the form of:

[0037] Lω T x i =λ i Dω T x i .

[0038] Preferably, in step 3, a basic unary function library is established to generate real network dynamics. and binary basic function library Including orthogonal basis functions, polynomial functions, exponential functions, fractional functions and Sigmoid, Relu.

[0039] Preferably, in step 4, reasoning the network dynamics equation from the basic function library based on the data space after dimensionality reduction specifically includes the following steps:

[0040] Step 41: The form of the kinetic equation is:

[0041]

[0042] Among them, Y i (t)≡(y i，1 (t), ..., y i，d (t)) T is the traffic data vector of base station numbered i, d is the dimension of the feature space after Laplace feature transformation, n is the total number of base station nodes, P(·) and Q(·,·) are the unary and binary basic function libraries respectively. Linear combinations of functions in ;

[0043] Step 42: Infer the coefficient ω of the optimal solution of the dynamic equation P 、ω Q :

[0044]

[0045] in, represents the Cartesian product, E d is the d-dimensional unit matrix.

[0046] Preferably, in step 42, the reasoning process of the kinetic equation specifically includes the following steps:

[0047] Step 421: Time-varying derivative of traffic at each node You can use y by the five-point difference method i (t) is approximately:

[0048]

[0049] Among them, δt is the time step;

[0050] Step 422: Solve the optimization problem with the regularization term to find the coefficient ω of the optimal solution of the inference dynamics equation. P 、ω Q :

[0051]

[0052] in, To control ω P 、ω Q Sparsity hyperparameters;

[0053] Step 423, delete one by one from small to large P 、ω Q The values ​​in , calculate the weighted Akaike Information Criterion AIC ω :

[0054]

[0055] When AIC ω When the decreasing trend of P 、ω Q As a final result;

[0056] Step 424: Substitute the coefficient vector finally obtained into step 42 and perform inner product operation with the function library as the final output dynamic equation form.

[0057] The beneficial effects of the present invention are as follows: (1) the present invention can effectively extract the topological properties of network nodes when facing large-scale graph data. The proposed input proxy module effectively preserves node information through matrix coding and has strong universality; (2) the scale of network nodes in real scenarios is tens of thousands, and the use of traditional base station traffic analysis methods will produce massive hyperparameters. The feature space dimensionality reduction module proposed by the present invention effectively reduces the dimension of the feature space while retaining most of the feature information, which is conducive to model fitting and reasoning, and greatly improves the computational efficiency; (3) the present invention uses the time-varying properties of the dynamic equation to capture the self-evolution and interaction effects of network nodes, and infers the equation from the basic function library, which can fully capture the intrinsic characteristics and potential laws of communication base station traffic. The parameters of the dynamic equation model are known and the interpretability is strong, which can achieve better accuracy for both short-term and long-term traffic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the system structure of the present invention.

[0059] Figure 2 Schematic diagram of flow information dimensionality reduction in the present invention.

[0060] Figure 3 It is a structural diagram of the basic function module in the present invention. DETAILED DESCRIPTION

[0061] like Figure 1 As shown, a communication base station traffic analysis system based on graph embedding technology and dynamic equation reasoning includes: a communication base station information acquisition module, an input agent module, a feature space dimensionality reduction module, a dynamic model inference module and a function library module;

[0062] The communication base station information collection module is composed of the communication base station's locator and traffic information detector to form an information collection network. The locator records the base station's spatial information, and the traffic information detector captures the direction and intensity of time-varying traffic.

[0063] The input agent module encodes the network data input by the communication base station information acquisition module into a matrix form and inputs it into the feature space dimensionality reduction module;

[0064] The feature space dimensionality reduction module fully extracts spatiotemporal features to achieve data dimensionality reduction, and inputs the low-dimensional data into the dynamic model inference module;

[0065] The dynamic model inference module searches for the most explanatory and concise function model in the function library module;

[0066] The library module stores the basic functions that may generate real network dynamics.

[0067] The goal of base station traffic analysis is to infer future traffic trends or predict traffic intensity at a certain time in the future based on currently observed base station network information (including location information, traffic direction information, and traffic intensity information). The base station traffic analysis method proposed in this invention, based on graph embedding technology and dynamic equation reasoning, includes the following steps:

[0068] S1, measured traffic data information of the communication base station network, the communication base station network is represented as a weighted acyclic graph D = (V, E, A), where V is the node set consisting of the communication base stations, and |V| = n, E is the edge set between nodes, and A is the adjacency matrix that measures the distance between nodes;

[0069] The location information of the base station is used to establish a weighted adjacency matrix A that represents the proximity of the nodes using a Gaussian kernel with a threshold:

[0070]

[0071] A ij Indicates base station v with number i i With base station v numbered j j The distance weight, ρ(v i , v j ) represents the base station v numbered i i With base station v numbered j j The distance between 2 is the variance of the distance between base stations, is the set distance threshold. According to the mapping of the Gaussian kernel, the farther the distance between the two base stations v i 、v j , its weight in the matrix A ij Therefore, during inference, the model will pay more attention to the traffic interaction between neighboring base stations.

[0072] S2, using graph embedding technology to reduce the dimensionality of the eigenvector space of the high-dimensional sparse adjacency matrix A, and project the time-varying traffic data into the low-dimensional feature space. Figure 2 The specific steps are as follows:

[0073] S2.1, the data dimension in large-scale communication base station networks is usually very high. Mapping data from high-dimensional space to low-dimensional space can reduce the high dimensionality and the large number of hyperparameters in the model. Using the mapping function F(·), each feature vector Convert to low-dimensional representation Make Y i (t) = F(X i (t)), M is the dimension of the original feature space, and d is the dimension of the transformed feature space;

[0074] S2.2, calculate the Laplace matrix L≡DA, L is a real symmetric matrix, and its semi-positive definiteness can be proved. The diagonal matrix D stores the degree of the base station node,

[0075] S2.3, in graph embedding, we usually use the first-order similarity to measure whether two nodes are similar. We use the edge weight A ij The similarity features of the nodes are recorded. The idea behind the Laplace eigenmap is that if two nodes are highly similar in the original network, then their embedding vectors should be close to each other in the low-dimensional space. Therefore, the optimization problem should be set so that nodes with higher similarity in the low-dimensional space are closer to each other, that is, satisfying:

[0076]

[0077] However, in order to prevent some extreme cases (such as all embedded vectors overlap, the optimization function Y * = 0, such a result is meaningless), we need to give a constraint condition so that all embedded vectors fill the low-dimensional space as much as possible So given the constraint Y T DY=E, where E is the unit matrix.

[0078] Next we choose the linear projection Y=X T w completes the mapping, transforming the feature space from down to At this time, the objective function becomes:

[0079]

[0080] The constraints become:

[0081] ω T X T DXw=E

[0082] The Lagrangian function is constructed as:

[0083]

[0084] Among them, Λ is a diagonal matrix, and the elements Λ on the diagonal ii =λ i is the Lagrange multiplier;

[0085] Solving the Lagrange equation to obtain the coefficient ω of the linear projection, the matrix form of the equation is:

[0086] Lω T X=-Dw T XΛ

[0087] The general form of the equation is:

[0088] Lω T x i =λ i Dω T x i

[0089] The coefficient ω can be obtained by referring to the solution method of the generalized eigenvalue problem.

[0090] S3, respectively establish a basic unary function library that can generate real network dynamics and binary basic function library Including basic functions such as orthogonal basis functions, polynomial functions, exponential functions, fractional functions, and common activation functions in other fields such as Sigmoid and Relu. Figure 3 shown.

[0091] S4, inferring the network dynamics equations from the basic function library based on the reduced-dimensional data space. The specific steps are as follows:

[0092] S4.1. In reality, the behavior of a complex network system arises not only from individual nodes but also from the dynamic interactions between them. We characterize the behavior of a complex network system using the following dynamics equation (Barzel et al., 2013, Universality in Network Dynamics):

[0093]

[0094] Among them, Y i (t)≡(y i,1 (t), ..., y i，d (t)) T is the traffic data vector of base station numbered i, d is the dimension of the feature space after Laplace feature transformation, n is the total number of base station nodes, P(·) and Q(·,·) are the unary and binary basic function libraries respectively. Linear combinations of functions in ;

[0095] S4.2, since the unary function P(·) that characterizes the action of a single node and the binary function Q(·,·) that characterizes the interaction between nodes are both from the basic function library, we will is considered as a function vector and the flow data vector is used as the input of each function. The inference problem can be reformulated as inferring the coefficients ω of the optimal solution of the dynamic equation P 、ω Q :

[0096]

[0097] in, denotes the Cartesian product, E d is a d-dimensional identity matrix.

[0098] For the derivative term, referring to the numerical solution of stochastic differential equations, the time-varying derivative of each node flow can be approximated by the five-point difference method with y i (t) as:

[0099]

[0100] where δt is the time step;

[0101] The optimization problem for solving the coefficients ω P , ω Q should be set so that the inferred model results are as close as possible to the true results in a time window [0, T]. In order to keep the parameters relatively stable, we introduce a regularization term for the coefficients ω P , ω Q in the optimization equation:

[0102]

[0103] where, is a hyperparameter that controls the sparsity of ω P , ω Q ;

[0104] Considering that some parameter values may be too small, contributing less to the interpretability of the model, while increasing the complexity of the model. Since the Akaike information criterion (AIC) can balance the complexity of the estimated model and the goodness of the model fitting data, we delete the values in ω P , ω Q one by one from small to large, and calculate the weighted Akaike information criterion AIC ω :

[0105]

[0106] The term with a larger ω is more likely to capture the potential dynamic characteristics, and adding a weight ω to the Akaike information criterion can amplify the influence of this effect. When the decreasing trend of AIC ω appears an inflection point, the current coefficient vector ω P , ω Q is taken as the final result.

[0107] After obtaining the final coefficient vector ω P , ω Q , the inner product operation of the coefficient vector and the function library is performed, which is the final output form of the dynamic equation.

Claims

1. A communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning, characterized in that: The steps include: Step 1: Measure the traffic data of the communication base station network and represent the communication base station network as a weighted acyclic graph D = (V, E, A), where V is the node set consisting of the communication base stations, E is the edge set between the nodes, and A is the adjacency matrix that measures the distance between the nodes. The calculation method of the weighted adjacency matrix A, which represents the node proximity, is: A ij Indicates base station v with number i i With base station v numbered j j The distance weight, ρ(v i ,v j ) represents the base station v numbered i i With base station v numbered j j The distance between 2 is the variance of the distance between base stations, is the set distance threshold; Step 2: Use graph embedding technology to reduce the dimensionality of the eigenvector space of the high-dimensional and sparse adjacency matrix A and project the time-varying traffic data into a low-dimensional feature space; Step 3: Build two basic function libraries; Step 4: Inferring the network dynamics equation from the basic function library based on the reduced-dimensional data space; specifically, the following steps are included: Step 41: The form of the kinetic equation is: Among them, Y i (t)≡(y i，1 (t), ..., y i，d (t)) T is the traffic data vector of base station numbered i, d is the dimension of the feature space after Laplace feature transformation, n is the total number of base station nodes, P(·) and Q(·,·) are the unary and binary basic function libraries respectively. Linear combinations of functions in ; Step 42: Infer the coefficient ω of the optimal solution of the dynamic equation P 、ω Q : in, represents the Cartesian product, E d is the d-dimensional unit matrix.

2. The communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning according to claim 1 is characterized in that: In step 2, graph embedding technology is used to reduce the dimensionality of the eigenvector space of the high-dimensional and sparse adjacency matrix A. Projecting the time-varying traffic data into the low-dimensional feature space specifically includes the following steps: Step 21: Use the mapping function F(·) to transform each eigenvector Convert to low-dimensional representation Make Y i (t) = F(X i (t)), M is the dimension of the original feature space, and d is the dimension of the transformed feature space; Step 22: Calculate the Laplace matrix L≡DA. The diagonal matrix D stores the degree of the base station node. Step 23: Given the constraint Y T DY=E, find the optimal function F(·) under the constraints:

3. The communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning according to claim 2 is characterized in that: In step 23, given the constraint Y T DY=E, find the optimal function F(·) under the constraints: After using linear projection for dimensionality reduction, the process of solving the coefficients includes the following steps: Step 231: Use Y=X T The linear projection of w will be down to Step 232: Use linear projection Y=X T w, the objective function becomes: Step 233: Under the constraint condition ω T X T When DXω=E, the Lagrangian function is constructed as: Among them, Λ is a diagonal matrix, and the elements Λ on the diagonal ii =λ i is the Lagrange multiplier; Step 234: Use the generalized eigenvalue method to solve the Lagrange equation to obtain the coefficient ω of the linear projection. The equation is in the form of: Lω T x i =λ i D T x i 。 4. The communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning according to claim 1 is characterized in that: In step 3, a basic unary function library is established to generate real network dynamics. and binary basic function library Including orthogonal basis functions, polynomial functions, exponential functions, fractional functions and Sigmoid, Relu.

5. The communication base station traffic analysis method based on graph embedding technology and dynamic equation reasoning according to claim 1 is characterized in that: In step 42, the reasoning process of the dynamic equation specifically includes the following steps: Step 421: Time-varying derivative of traffic at each node You can use y by the five-point difference method i (t) is approximately: Among them, δt is the time step; Step 422: Solve the optimization problem with the regularization term to find the coefficient ω of the optimal solution of the dynamics equation P 、ω Q : Where λ is the control ω P 、ω Q Sparsity hyperparameters; Step 423, delete one by one from small to large P 、ω Q The values ​​in , calculate the weighted Akaike Information Criterion AIC ω : When AIC ω When the decreasing trend of P 、ω Q As a final result; Step 424: Substitute the coefficient vector finally obtained into step 42 and perform inner product operation with the function library as the final output dynamic equation form.

Images