A wireless communication network optimization method based on knowledge graph causal effect estimation

Abstract

The application discloses a wireless communication network parameter tuning method based on knowledge graph causal effect estimation. The method comprises the following steps: starting from wireless communication network parameter measurement data, constructing a corresponding local knowledge graph according to the endogenous factors of the wireless communication network protocol, separating the research object, the intervention variable, the covariate and the observation result from the measurement data, constructing a random test using the observation data, evaluating the influence effect of the intervention variable on the network performance index using the method of counterfactual reasoning, generalizing to the entire graph, and obtaining the influence effect of each adjustable parameter on each network performance index, thereby giving the direction of wireless communication network performance adjustment and giving the network optimization adjustment options on the basis of the existing network. The system comprises a knowledge graph construction module, an environment interaction module and a causal reasoning module. The application optimizes the wireless communication network parameters under the condition of limited computing resources and labor costs, starting from the global optimization of the wireless communication network.

Description

Technical Field

[0001] This invention relates to a method and system for parameter tuning of wireless communication networks based on causal effect estimation, belonging to the fields of wireless communication, knowledge graphs, and causal reasoning. Background Technology

[0002] In recent years, with the large-scale deployment of 5G mobile communication networks, the number of wireless communication network devices has surged. The massive amounts of data generated by these network devices have provided an opportunity to use artificial intelligence methods to analyze the internal parameter relationships of communication networks and improve network performance.

[0003] Wireless communication networks involve numerous network elements, and optimizing them requires considering many adjustable parameters. Traditional network optimization methods involve modeling the entire wireless network environment after measuring environmental parameters (such as building and vegetation obstruction) in the field. A series of formulas from communication theory are then used to list the constraints on the adjustable parameters, and the optimal solution is obtained by solving the optimization problem. This process is a static engineering approach; these formulas and communication theories do not need to be repeated in every engineering calculation. However, current wireless communication networks are widely distributed, and the wireless communication environment is constantly changing. This model-dependent optimization method requires significant human and material resources, and the optimization results are often delayed.

[0004] Classical neural network algorithms analyze the correlations between data, typically for prediction purposes, and require the data to be labeled. However, wireless communication network parameter data lacks relevant labels. Furthermore, wireless communication fields have explicit physical meanings, and the relationships between these fields cannot be utilized by classical neural network algorithms. Therefore, wireless communication network data cannot be directly used with neural network algorithms. Summary of the Invention

[0005] Purpose of the invention: This invention aims to address the complex problem of parameter tuning in wireless local area networks (WLANs) by proposing a method and system for wireless communication network parameter tuning based on knowledge graph causal effect estimation.

[0006] Based on the wireless communication network parameter tuning system of the present invention, the present invention proposes a method for tuning wireless communication network parameters based on knowledge graph causal effect estimation, including the following steps:

[0007] Step 1: Based on the resource scheduling process in the downlink communication flow of wireless communication, collect network parameters, resource scheduling parameters, and network performance indicator parameters as entity nodes in the knowledge graph to construct a local knowledge graph for downlink communication in wireless communication. Downlink throughput is affected by four factors: spectral efficiency (MCS), bandwidth (RB), bandwidth grant (DL grant), and bit error rate (BLER). Relationships between parameters with definite formula constraints are defined as causal relationships, and the corresponding relationship direction in the knowledge graph is from cause to effect. Relationships between parameters that influence each other but do not have definite formula constraints are defined as association relationships, and the corresponding relationship direction in the knowledge graph is along the chronological order of the downlink scheduling flow.

[0008] Step 2: The network parameters and resource scheduling parameters that can be adjusted by humans are called adjustable parameters, which are used as the selection range of intervention variables in causal inference. The network performance indicator parameters of interest are used as the selection range of outcome variables in causal inference. The parameters that cannot be adjusted by humans are called unadjustable parameters, which are used as the selection range of covariates in causal inference.

[0009] Step 3: In each round of causal inference, select the parameters that need to be adjusted from the adjustable parameters as the intervention variables in this round of causal inference, select the network performance indicator variables of interest as the outcome variables, find the corresponding covariates according to the relationships in the knowledge graph, use the causal effect estimation algorithm to estimate the causal relationship between the intervention variables and the outcome variables, and output the impact of different values ​​of the intervention variables on the outcome variables under the condition of covariate stratification, thereby obtaining the adjustment direction of the adjustable variables;

[0010] Furthermore, step 1 specifically includes:

[0011] Step 1.1: According to the wireless communication downlink resource scheduling process, obtain the network parameters, resource scheduling parameters and network performance indicator parameters involved in the scheduling process, and use these parameters as nodes of the knowledge graph;

[0012] The network parameters include CQI (Channel Quality Indicator), CR (Code Rate), RI (Rank Indication), RS (Reference Signal Received Power), RSSI (Received Signal Strength Indicator), RSRQ (Reference Signal Received Quality), and SINR (Signal to Interference plus Noise Ratio). Among these, RSRP and SINR are radio parameters, while RI and CQI are feedback parameters.

[0013] The resource scheduling parameters include PRB (Physical Resource Block), MCS (Modulation and Coding Scheme), TB Size (Transport Block Size), HARQ parameters, BLER (Block Error Rate), ACK (Acknowledgment Frame), NACK (Back Off Frame), MCS modulation scheme, 5G uplink and downlink slot rates, 5G reference time slot configuration, 5G uplink and downlink switching cycle mode, number of 5G downlink PRBs, and number of 5G downlink slots, etc.

[0014] The network performance indicators include: 5G downlink physical layer average rate, 5G downlink MAC layer average rate, 5G downlink RLC layer average rate, 5G downlink PDCP layer average rate, dual-connectivity downlink physical layer throughput, dual-connectivity downlink MAC layer average rate, dual-connectivity downlink RLC layer throughput, and dual-connectivity downlink PDCP layer throughput.

[0015] Step 1.2: Based on the interaction between HARQ and BLER parameters in the radio parameters, feedback parameters, and resource scheduling parameters during downlink resource scheduling, the causal and correlation relationships between the parameters are obtained. Then, for the existing parameters, their causal variables are found from the communication protocol. For example, based on the mapping relationship between reference signals:

[0016]

[0017]

[0018]

[0019] It is possible to determine the causal relationship between reference signals. Here, RSSI represents the linear average of the signal power received on all OFDM symbols containing the reference signal within the measurement bandwidth, including all signal quantities such as signals from the current cell and neighboring cells in the same frequency band at this location, adjacent channel interference, and thermal noise. Inerference represents the interference power. Channel Quality Indicator (CQI) is the cause of Signal-to-Interference-plus-Noise Ratio (SINR). Reference Signal Received Strength (RSRP) is the cause of Reference Signal Received Quality (RSRQ) and SINR, etc. Furthermore, based on the theoretical calculation formula for downlink peak rate:

[0020] R peak =V layer ·Q m ·R max ·N PRB ·(1-OH)·N sys

[0021] Among them, R peak V represents the peak rate. layer Indicates the number of MIMO layers, Q m Indicates modulation order, R max Indicates the coding rate, N PRB Indicates the number of PRBs, OH represents the resource consumption percentage, and N sys The sign indicates the number of signs. The parameter on the right side of the equals sign in the formula has a causal relationship with the peak rate and is the cause variable of the peak rate.

[0022] Furthermore, step 3 specifically includes:

[0023] Step 3.1: Based on the wireless communication downlink knowledge graph constructed in Step 1, find the causes of intervention variables, outcome variables, and covariates. For intervention variables without direct causes, use a random forest classifier to assign a classifier function causal model. For outcome variables and covariates without direct causes, specify an additive noise model of random forest regression as the noise function and empirical distribution.

[0024] Step 3.2: Before estimating the causal effect, first observe the statistical analysis of the covariates and divide the values ​​of the covariates into several discrete points to study the impact of the intervention variable on the outcome variable under different covariate conditions. In wireless communication, this means studying the impact of adjustable parameters on network performance indicator parameters under different interference conditions.

[0025] Step 3.3: Calculate the conditional average treatment effect (CATE) of the intervention on the outcome variable. To do this, the intervention variable in the fitted causal plot is randomly intervened, samples are extracted from the intervention distribution, and the intervention effect for each subsample is calculated using the stratified average causal effect. This includes the following steps:

[0026] Step 3.3.1: Model the downlink knowledge graph of wireless communication using a causal model. The causal model includes a causal graph and a structural assumption, namely, assuming a causal relationship between the intervention variable T and the outcome variable Y, and using the causal graph containing this directed edge as the model. The causal graph is a directed acyclic graph containing entities and relations in the knowledge graph, where the nodes of the causal graph correspond to entities in the knowledge graph, and the directed edges of the causal graph correspond to relations in the knowledge graph. The structural assumption is that a causal relationship exists between the intervention variable and the outcome variable.

[0027] Step 3.3.2: Determine whether the desired effect under the causal model can be estimated. If the intervention variable and the outcome variable do not have a common cause, a causal effect can be identified. If the intervention variable and the outcome variable have a common cause and there are enough observed variables to block all backdoor paths, a causal effect can also be identified.

[0028] Step 3.3.3: Estimate the effect using a statistical estimator. Here, CATE is used for estimation. For each value l of the covariate L, for both binary intervention (T = 0, 1) and binary outcome (Y = 0, 1),

[0029] CATE=E[Y(T=1)|L=l]-E[Y(T=0)|L=l]

[0030] In this application scenario, the intervention variable (adjustable communication parameter) takes discrete multiple values. Therefore, the CATE between each pair of values ​​is calculated, and the CATE at each T=t can be obtained. Then, the average is taken. The CATE between each pair of values ​​obtained in this way is the effect of the change of the intervention variable on the outcome variable.

[0031] Step 3.3.4: refute the obtained estimate through robustness checks and sensitivity analysis;

[0032] Step 3.3.5: The above estimates are used as the causal effect of the intervention variable on the outcome variable, that is, the left and right estimates of the network performance indicator parameter by changing the (directional) adjustable parameter.

[0033] The present invention provides a wireless communication network parameter optimization system based on causal reasoning, comprising a knowledge graph construction module, an environment interaction module, and a causal reasoning module;

[0034] Knowledge Graph Construction Module: Based on the wireless communication downlink resource scheduling process and wireless communication protocol, this module constructs a wireless communication downlink knowledge graph. The resulting causal graph is then sent to the environment interaction module and the causal reasoning module. The causal graph includes nodes and the relationships between them, including causal and associative relationships.

[0035] The environment interaction module cleans and organizes the network parameters, resource scheduling parameters, and network performance indicator parameters. These parameters are then divided into adjustable parameters, non-adjustable parameters, and network performance indicator parameters according to the constructed knowledge graph. They are then sent to the causal reasoning module as alternative intervention variables, covariates, and outcome variables in the causal reasoning process.

[0036] Causal Reasoning Module: This module uses the causal graph output by the knowledge graph construction module and the intervention variables, outcome variables, and covariates output by the environment interaction module to perform causal reasoning. It outputs the impact of different values ​​of the intervention variable on the outcome variable, i.e., the average causal effect, and provides the direction for adjusting the intervention variable and the expected effect.

[0037] The advantages of this invention compared to existing technologies lie in the introduction of knowledge graphs and causal reasoning. The knowledge graph of endogenous factors in wireless communication represents the relationships between numerous communication protocols and communication data fields using a directed acyclic graph, facilitating a holistic understanding of the relationships between various parameters of the wireless communication network. By utilizing causal reasoning and fully leveraging the vast amounts of data generated by the wireless communication network, combined with expert knowledge from the knowledge graph of endogenous factors in wireless communication, and under the premise of automated measurement of network node data, the impact of adjustable parameters on network performance indicators can be inferred in real time using current measurement data. This provides the direction for adjusting adjustable parameters, enabling real-time control of the wireless communication network status with reduced manual calculations, resulting in better network optimization.

[0038] This invention can provide an adjustment scheme to improve the downlink throughput of wireless communication networks after a single calculation, targeting scenarios of optimizing downlink throughput in wireless communication. Attached Figure Description

[0039] Figure 1 This is a diagram of the causal reasoning wireless communication network parameter optimization architecture according to an embodiment of the present invention;

[0040] Figure 2 This is a detailed flowchart of the knowledge graph construction module according to an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the main part of the wireless communication downlink knowledge graph used in the embodiments of the present invention;

[0042] Figure 4This is a detailed flowchart of the environment interaction module in an embodiment of the present invention;

[0043] Figure 5 This is a detailed flowchart of the causal reasoning module in an embodiment of the present invention;

[0044] Figure 6 This is a causal effect diagram output by the causal reasoning of the embodiment. Detailed Implementation

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] This embodiment provides a wireless communication network optimization method based on knowledge graph causal effect estimation, which solves the problem of complex calculations for downlink throughput tuning in traditional wireless networks. Figure 1 The present invention illustrates a causal inference wireless communication network parameter adjustment architecture diagram, including:

[0047] Knowledge graph construction module: Constructs a wireless communication downlink knowledge graph based on the wireless communication downlink resource scheduling process and wireless communication protocol. The module outputs the causal graph of the wireless communication downlink knowledge graph and sends it to the causal reasoning module.

[0048] The environmental interaction module takes the measured wireless communication network parameters, cleans and organizes them, and then divides them into intervention variables, outcome variables, and covariates according to the constructed knowledge graph before sending them to the causal reasoning module.

[0049] Causal Reasoning Module: This module uses the causal graph output by the knowledge graph construction module and the intervention variables, outcome variables, and covariates output by the environment interaction module to perform causal reasoning. It outputs the impact of different values ​​of the intervention variable on the outcome variable, i.e., the average causal effect, and provides the direction for adjusting the intervention variable and the expected effect.

[0050] Figure 2 The flowchart shows the detailed process of the knowledge graph construction module, including:

[0051] Step S101: According to the downlink communication protocol process of wireless communication, find the relevant communication parameters as entity nodes of the knowledge graph, and divide these entity nodes into adjustable parameters, non-adjustable parameters and network performance indicator parameters.

[0052] Step S102: According to the downlink communication process of wireless communication, find the relationship between entities based on the calculation and constraint relationship, and correspond to the directed edges on the knowledge graph.

[0053] Step S103: Construct a wireless communication downlink knowledge graph based on the aforementioned entity nodes and directed edges. A partial view of the knowledge graph is shown below. Figure 3Among them, the adjustable parameters are uplink and downlink slot rate, downlink PRB number and average MCS, the network performance index parameter is physical layer downlink throughput, and the non-adjustable parameter is average TB size; this knowledge graph is sent to the causal inference module in the form of a causal graph.

[0054] Figure 4 The flowchart showing the detailed process of the environment interaction module is presented, including:

[0055] Step S201: Organize the collected data, remove missing data, and normalize continuous data.

[0056] Step S202: For parameters whose causes cannot be found in the knowledge graph, observe their data distribution. For parameters without obvious distribution patterns, use the empirical distribution as the distribution of their causes.

[0057] Step S203: Classify the adjustable parameter data in the sorted data as intervention variable data, classify the network performance indicator variable data as outcome variable data, classify the remaining data as covariate data, and send them to the causal inference module.

[0058] Figure 5 A detailed flowchart of causal reasoning is shown, including:

[0059] Step S301: For each adjustable parameter, find the covariates between the adjustable parameter and the network performance indicator variable, that is, the variables that affect the value of the adjustable variable and the variables that affect both variables, as well as the causal variables of these variables, as input variables for causal inference.

[0060] Step S302: Input the intervention variable, outcome variable, covariate, and causal graph into the Dowhy algorithm.

[0061] Step S303: Obtain the effect of the intervention variable on the outcome variable under the condition of covariates, and make a network parameter adjustment plan based on this effect.

[0062] The reasoning results of the above algorithm on the impact of an adjustable communication parameter on network performance indicators are as follows: Figure 6 As shown, under the condition of the covariate nr_ul_dl_slot_ratio (uplink and downlink slot rates), when this variable has a large value (80%-100%), the intervention variable prb_num_dl_s (number of downlink PRBs) has a significant impact on the result variable nr_phy_throughput_dl (physical layer downlink throughput). Therefore, in this case, increasing the number of downlink PRBs helps to improve the physical layer downlink throughput.

[0063] This invention visualizes the parameter relationships in a wireless communication network using a knowledge graph, eliminating the need for manual calculations. It employs causal reasoning to reveal the causal relationships between network parameters under the current network state. Compared to traditional wireless network optimization schemes, the proposed method uses existing network measurement data to provide adjustment directions for adjustable parameters based on network performance indicators, avoiding the need to calculate complex formulas for each parameter adjustment. As the network environment changes, targeted adjustments can be made based on real-time updated measurement data. While directly applying neural networks to wireless communication data is still challenging, using knowledge graphs, incorporating expert knowledge, and leveraging the inherent relationships within the wireless communication knowledge graph contributes to fully utilizing the massive amounts of data generated by wireless communication networks.

[0064] The present invention has been disclosed above with reference to preferred embodiments, but is not intended to limit the technical solutions in the related field. Those skilled in the art can make extensions and transfers based on the present invention, and all work that does not constitute an inventive contribution is covered within the scope of the claims of the present invention.

Claims

1. A wireless communication network optimization method based on knowledge graph causal effect estimation, characterized in that, Includes the following steps: Step 1: Based on the resource scheduling process in the downlink communication flow of wireless communication, collect network parameters, resource scheduling parameters, and network performance indicator parameters as entity nodes in the knowledge graph, construct a local knowledge graph for downlink communication in wireless communication, define the relationship between parameters with definite formula constraints as causal relationship, and the corresponding relationship direction in the knowledge graph is from cause to effect; define the relationship between parameters that influence each other but have no definite formula constraints as association relationship, and the corresponding relationship direction in the knowledge graph is along the sequential direction of the downlink scheduling flow; Step 2: The network parameters and resource scheduling parameters that can be adjusted manually are called adjustable parameters, which serve as the range of intervention variables to be selected in causal inference; Network performance indicator parameters are used as the range of outcome variables in causal inference; Parameters that cannot be adjusted artificially are called unadjustable parameters, and are used as the range of covariates to be selected in causal inference; Step 3: In each round of causal inference, select the parameters that need to be adjusted from the adjustable parameters as the intervention variables in this round of causal inference, select the network performance indicator variables of interest as the outcome variables, find the corresponding covariates according to the relationships in the knowledge graph, use the causal effect estimation algorithm to estimate the causal relationship between the intervention variables and the outcome variables, and output the impact of different values ​​of the intervention variables on the outcome variables under the condition of covariate stratification, thereby obtaining the adjustment direction of the adjustable variables; Step 3 specifically includes: Step 3.1: Based on the wireless communication downlink knowledge graph constructed in Step 1, find the causal variables of the intervention variables, outcome variables, and covariates. For intervention variables without direct causes, use a random forest classifier to assign a classifier function causal model. For outcome variables and covariates without direct causes, specify an additive noise model of random forest regression as the noise function and empirical distribution. Step 3.2: First, observe the statistical analysis of the covariates and divide the values ​​of the covariates into several discrete points to study the impact of the intervention variable on the outcome variable under different covariate conditions. In wireless communication, this means studying the impact of adjustable parameters on network performance indicator parameters under different interference conditions. Step 3.3: Calculate the conditionally averaged intervention effect (CATE) on the outcome variable. To do this, randomize the intervention variable in the fitted causal plot, extract samples from the intervention distribution, and calculate the intervention effect for each subsample using the stratified average causal effect. This includes the following steps: Step 3.3.1: Model the data using causal diagrams and structural hypotheses, i.e., hypothesize intervention variables. and outcome variables There is a causal relationship between them, so we will use the causal graph containing this directed edge as the model; Step 3.3.2: Determine whether the desired effect can be estimated under the causal model. If the intervention variable and the outcome variable do not have a common cause, then identify the causal effect. If the intervention variable and the outcome variable have a common cause and the observed variable can block all backdoor paths, then identify the causal effect. Step 3.3.3: Estimate the effects using a statistical estimator; here, CATE is used for estimation of the covariates. Each value For binary interventions and binary outcomes, Since the intervention variable takes discrete multiple values, the CATE between each pair of values ​​is calculated to obtain each... The CATE values ​​are calculated and then averaged. The CATE values ​​between any two values ​​represent the effect of the change in the intervention variable on the outcome variable. Step 3.3.4: refute the obtained estimate through robustness checks and sensitivity analysis; Step 3.3.5: The above estimates are used as the causal effect of the intervention variable on the outcome variable, that is, the left and right estimates of the network performance indicator parameter by changing the adjustable parameter.

2. The wireless communication network optimization method based on knowledge graph causal effect estimation according to claim 1, characterized in that, Step 1 is as follows: Step 1.1: According to the wireless communication downlink resource scheduling process, obtain the network parameters, resource scheduling parameters and network performance indicator parameters involved in the scheduling process, and use these parameters as nodes of the knowledge graph; The network parameters include Channel Quality Indicator (CQI), Code Rate (CR), Rank (RI), Reference Signal (RS): Reference Signal Received Strength (RSRP), Received Signal Strength Indicator (RSSI), Reference Signal Received Quality (RSRQ), and Signal-to-Interference-plus-Noise Ratio (SINR). The resource scheduling parameters include Physical Resource Block (PRB), Modulation and Coding Scheme (MCS), Transport Block Size (TB Size), HARQ parameters, Decoding Error Rate (BLER), Acknowledgment Frame (ACK), Back Off Frame (NACK), Modulation Scheme (MCS), 5G Uplink and Downlink Slot Rates, 5G Reference Time Slot Configuration, 5G Uplink and Downlink Switching Cycle Mode, Number of 5G Downlink PRBs, and Number of 5G Downlink Slots. The network performance indicators include the average rate of the 5G downlink physical layer, the average rate of the 5G downlink MAC layer, the average rate of the 5G downlink RLC layer, the average rate of the 5G downlink PDCP layer, the downlink physical layer throughput of dual connectivity, the downlink MAC layer throughput of dual connectivity, the downlink RLC layer throughput of dual connectivity, and the downlink PDCP layer throughput of dual connectivity. Step 1.2: Based on the interaction between the wireless parameters and feedback parameters in the network parameters and the HARQ parameters and BLER parameters in the resource scheduling parameters, obtain the causal relationship and correlation between the parameters. For the existing parameters, find their causal variables from the communication protocol.

3. The wireless communication network optimization method based on knowledge graph causal effect estimation according to claim 2, characterized in that, The adjustable parameters in step 2 include modulation scheme (MCS), 5G uplink and downlink slot rate, 5G reference time slot configuration, 5G uplink and downlink switching cycle mode, number of 5G downlink PRBs, number of 5G downlink slots, and RI.

4. A system for implementing the wireless communication network optimization method based on knowledge graph causal effect estimation as described in claim 1, characterized in that, It includes a knowledge graph construction module, an environment interaction module, and a causal reasoning module; Knowledge graph construction module: Constructs a wireless communication downlink knowledge graph based on the wireless communication downlink resource scheduling process and wireless communication protocol. The knowledge graph construction module outputs the causal graph of the wireless communication downlink knowledge graph to the environment interaction module and the causal reasoning module. The environment interaction module categorizes wireless communication network parameter data into intervention variables, outcome variables, and covariates according to the constructed knowledge graph, and inputs them into the causal reasoning module. Causal Reasoning Module: This module uses the causal graph output by the knowledge graph construction module and the intervention variables, outcome variables, and covariates output by the environment interaction module to perform causal reasoning. It outputs the impact of different values ​​of the intervention variable on the outcome variable, i.e., the average causal effect, and provides the direction for adjusting the intervention variable and the expected effect.

Images