Causal discovery method and system for wireless network performance indexes

Abstract

The invention relates to the technical field of artificial intelligence and causal relationship discovery, and provides a wireless network performance index-oriented causal discovery method and system, and the method comprises the following steps: collecting performance index data of a plurality of observation variables of a wireless network, and forming wireless network performance index data; calculating a causal difference between the observation variables of the wireless network performance index data by using the high-order accumulation amount; searching root nodes of a causal structure corresponding to the wireless network performance index data step by step based on the causal difference; taking the searched sequence of the root nodes as a causal sequence, and recovering the causal structure of the wireless network performance index data according to the causal sequence; according to the causal discovery method, the causal relationship can be accurately identified, and the calculation complexity is low.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and causal relationship discovery, and more specifically, to a causal relationship discovery method and system for wireless network performance indicators. Background Art

[0002] Generally speaking, the analysis of wireless network performance is a complex and critical task, involving the evaluation of multiple network indicators, such as signaling channel availability, signaling channel congestion, channel call drop rate, voice channel availability, voice channel congestion, voice channel call drop rate, handover success rate, dual-band handover success rate, wireless channel utilization, and uplink and downlink voice quality. These indicators, from the overall to the local, run through all periods of wireless network operation and are the main means of evaluating network performance. However, identifying and improving the causal relationship between these performance indicators is a key challenge to improving wireless network performance. In certain situations, inferring these causal relationships can help us find ways to optimize wireless network performance.

[0003] Several methods have been proposed in the prior art to identify the causal relationship between two observed variables by discovering asymmetry. Independence test is a typical tool used to determine the (in)dependence between two observed variables. In order to determine the causal direction between two adjacent variables, conditional independence tests (CITs) and V structure identification conditions can be used. Hilbert-Schmidt independence criterion (HSIC) test and kernel-based conditional independence (KCI) test are two typical CIT methods in linear non-Gaussian or nonlinear cases. They use kernel-based techniques to measure the dependence between variables in the reproducing kernel Hilbert space (RKHS). This method may introduce a lot of uncertainty, resulting in the inclusion of redundant information when characterizing data features. In this case, their performance relies on large sample sizes, and the computational complexity is proportional to the cube of the sample size.

[0004] The prior art also proposes some methods for estimating linear non-Gaussian acyclic models. For example, one method is based on the independent component analysis (ICA) technique, which estimates a mixing matrix showing how independent components affect the observed variables. The mixing matrix is ​​then converted into a causal strength matrix. However, ICA may fall into a local optimum and produce an unreasonable solution. The prior art also proposes a method based on independent noise conditions, using the independence between the residuals obtained after regression and the assumed dependent variable. For example, the DirectLiNGAM algorithm identifies the variable with the greatest independence as an exogenous variable, then eliminates the influence of the identified exogenous variable, and continues to search for the variable with the greatest independence to reconstruct the causal graph.

[0005] These methods mainly use the least squares method for regression operations. However, the least squares regression method has limitations when processing non-Gaussian data, often resulting in inaccurate residuals, which in turn affects the subsequent independence test results. This leads to inaccurate identification of causal relationships of noisy heteroscedastic variables when processing wireless network performance indicator data, high computational complexity, or poor performance when processing noisy non-Gaussian heteroscedastic variable data. Summary of the invention

[0006] In order to overcome the defects of inaccurate causal relationship identification and calculation in the above-mentioned prior art, the present invention provides a causal discovery method for wireless network performance indicators with accurate causal relationship identification and low calculation complexity.

[0007] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0008] Collecting performance indicator data of several observation variables of the wireless network to form wireless network performance indicator data;

[0009] Calculating causal differences between observed variables of the wireless network performance indicator data using high-order accumulation quantities;

[0010] Finding the root node of the causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference;

[0011] The order in which the root nodes are found is regarded as the causal order, and the causal structure of the wireless network performance indicator data is restored according to the causal order.

[0012] The present invention also proposes a causal discovery system for wireless network performance indicators for implementing the above-mentioned causal discovery method for wireless network performance indicators, the system comprising:

[0013] A collection module, used to collect performance indicator data of several observed variables of the wireless network to form wireless network performance indicator data;

[0014] A causal difference calculation module, used to calculate the causal difference between the observed variables of the wireless network performance indicator data using high-order accumulation;

[0015] A root node search module, used to search for a root node of a causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference;

[0016] The causal structure recovery module is used to regard the order in which the root nodes are found as the causal order, and to recover the causal structure of the wireless network performance indicator data according to the causal order.

[0017] The present invention also proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the above-mentioned causal discovery method for wireless network performance indicators.

[0018] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0019] The present invention uses high-order accumulation to calculate the causal difference between the observed variables of the wireless network performance indicator data, and searches for the root node of the causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference, regards the order in which the root nodes are found as the causal order, and restores the causal structure of the wireless network performance indicator data according to the causal order. It uses the characteristic that the causal difference between the root node and the leaf node has the consistency of causal order, effectively handles the noise heteroscedasticity problem that may exist in the data, can accurately identify the causal relationship and has low computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flowchart of the causal discovery method for wireless network performance indicators proposed in Example 1;

[0021] Figure 2 A schematic diagram of the process of searching for a root node level by level proposed in Example 1;

[0022] Figure 3 This is an overall framework diagram of a causal discovery system for wireless network performance indicators proposed in Example 3. DETAILED DESCRIPTION

[0023] The accompanying drawings are only used for illustrative purposes and should not be construed as limiting the present embodiment;

[0024] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0025] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0026] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0027] Example 1

[0028] This embodiment proposes a causal discovery method for wireless network performance indicators. Figure 1 A schematic diagram of the process of the causal discovery method for wireless network performance indicators of this embodiment;

[0029] A causal discovery method for wireless network performance indicators proposed in this embodiment includes the following steps:

[0030] S1: collecting performance indicator data of several observed variables of the wireless network to form wireless network performance indicator data;

[0031] S2: using high-order accumulation, calculating the causal difference between the observed variables of the wireless network performance indicator data;

[0032] S3: searching step by step for a root node of a causal structure corresponding to the wireless network performance indicator data based on the causal difference;

[0033] S4: Considering the order in which the root nodes are found as the causal order, and restoring the causal structure of the wireless network performance indicator data according to the causal order.

[0034] In the specific implementation process, high-order accumulation is used to calculate the causal difference between the observed variables of the wireless network performance indicator data, and the root node of the causal structure corresponding to the wireless network performance indicator data is searched step by step based on the causal difference. The order in which the root nodes are found is regarded as the causal order, and the causal structure of the wireless network performance indicator data is restored according to the causal order. The causal difference between the root node and the leaf node has the characteristic of causal order consistency, which effectively handles the noise heteroscedasticity problem that may exist in the data, can accurately identify the causal relationship and has low computational complexity.

[0035] In an optional embodiment, the expression of the wireless network performance indicator data includes:

[0036] X=[x1,x2,…,x i ,…,x N ]

[0037] In the formula, x i represents the performance indicator data of the i-th observed variable of the wireless network, and N represents the total number of observed variables;

[0038] Before using the high-order accumulation to calculate the causal difference between the observed variables of the wireless network performance indicator data, the wireless network performance indicator data is standardized, and the expression of the standardized processing includes:

[0039]

[0040] In the formula, represents the normalized x i ,mean(x i ) represents x i The mean of .

[0041] In an optional embodiment, the expression for calculating the causal difference between the observed variables of the wireless network performance indicator data using the high-order accumulation includes:

[0042] τ ij =|C4(x i )C 1,3 (x i ,x j )-C 2,2 (x i ,x j )C 3,1 (x i ,x j )|

[0043] C4(x i )=Cum(x i ,x i ,x i ,x i )

[0044] C 1,3 (x i ,x j )=Cum(x i ,x j ,x j ,x j )

[0045] C 2,2 (x i ,x j )=Cum(x i ,x i ,x j ,x j )

[0046] C 3,1 (x i ,x j )=Cum(x i ,x i ,x i ,x j )

[0047] In the formula, τ io Represents the i-th observed variable x of the wireless network performance indicator data i and the jth observed variable x j ; Cum(·) represents the high-order cumulant function.

[0048] In an optional embodiment, the step of searching step by step for a root node of a causal structure corresponding to the wireless network performance indicator data based on the causal difference comprises:

[0049] S1: For each observed variable in the wireless network performance indicator data, calculate the sum of the causal differences between it and other observed variables to obtain the sum of the causal differences corresponding to each observed variable;

[0050] S2: The causal difference and the smallest observed variable are regarded as the root node x of the causal structure corresponding to the wireless network performance indicator data m ;

[0051] S3: Eliminate the root node x from the wireless network performance indicator data m The impact on other observed variables, and the residual matrix R is obtained;

[0052] S4: Taking each observed variable in the residual matrix R as the operation object, iteratively execute steps S1 to S3, search for the root node step by step, and record the order in which the root nodes are found. When all observed variables are identified as root nodes, stop the iteration and obtain the causal order of the wireless network performance indicator data.

[0053] In this optional embodiment, Figure 2 A schematic diagram of the process of searching for a root node level by level proposed in this embodiment; Figure 2 The steps of searching step by step for the root node of the causal structure corresponding to the wireless network performance indicator data based on the causal difference are shown.

[0054] In an optional embodiment, the expression of the causal difference sum includes:

[0055]

[0056] Where, T τ (x i ; U) represents x i The causal difference sum, U represents the subscript set of variables in the feature set X, and U\i represents removing the i-th variable x from the subscript set U. i .

[0057] In an optional embodiment, the root node x m The expressions include:

[0058]

[0059] Where, T τ (x i ; U) represents x i The causal difference and.

[0060] In an optional embodiment, the expression of the residual matrix R includes:

[0061] R=[∈1,∈2,…,∈ m-1 ,∈ m+1 ,…,∈N ]

[0062] x pi =b mi x m -∈ i

[0063] In the formula, x m represents the root node; x pi represents the observed variable corresponding to the non-root node; b mi represents the regression coefficient; ∈ m-1 Represents the observed variable x m-1 The corresponding residual also indicates that the observed variable x is eliminated m-1 After the influence of m-1 The remaining part; ∈ i Represents x pi The corresponding residual; N represents the total number of observed variables.

[0064] In an optional embodiment, the step of restoring the causal structure of the wireless network performance indicator data according to the causal order includes:

[0065] Construct a lower triangular matrix B using the causal order, and let the element of the i-th row and j-th column of the lower triangular matrix B be B ij , if i≤j, then B ij =0, indicating x i x j There is no direct causal influence;

[0066] Otherwise, B ij Represents the observed variable x in the causal sequence i For the observed variable x j The direct causal influence of j The causal parent node is {x1,x2,…,x j-1}, then B ij The expressions include:

[0067] B ij =b ji

[0068]

[0069] In the formula, ∈ j Represents x j The corresponding residual.

[0070] Example 2

[0071] This embodiment is based on the causal discovery method for wireless network performance indicators proposed in Embodiment 1, and proposes the following specific implementation examples:

[0072] The present invention provides a causal relationship discovery method based on high-order accumulation, comprising:

[0073] S1: Set the observation data matrix X and standardize the data;

[0074] S2: For the standardized data, calculate the causal difference between each variable and other variables, determine the causal difference between each variable, and then use the summation method to sum the causal differences between each variable and other variables to obtain the sum of the causal differences of each variable;

[0075] S3: According to the obtained causal difference sum of each variable, determine the variable with the smallest causal difference sum as the root node;

[0076] S4: Other variables are regressed on the root node respectively to calculate the residual matrix R.

[0077] S5: Replace the data matrix X with the residual matrix R, and repeat steps 2 to 4 for R until all root nodes are found in sequence. The order in which the root nodes are found is the causal order, and the causal structure is restored according to the causal order.

[0078] S6: Based on the directed acyclic graph model found by the causal relationship discovery algorithm, the causal relationship results between the wireless network performance indicator data are obtained and fed back to the wireless network maintenance personnel, etc., to provide assistance for scientific discovery in wireless network performance optimization problems.

[0079] The technical solution for causal discovery of wireless network performance indicator data is a method for linear non-Gaussian causal structure discovery, which identifies the causal structure of two variables by establishing statistics (causal difference) based on high-order accumulation, and proves that the method of the present application can be extended to multivariate situations, and proves the advantages of the method of the present application over other methods in the case of noise heteroscedasticity. It is characterized by considering the high-order information of the observed variables, combining the idea of ​​multivariate direct lingam algorithm, and adopting the idea of ​​identifying the root node of causal order to transform the difficult problem into an easy-to-solve problem, including the following steps:

[0080] Set the wireless network performance indicator data X = [x1, x2, ... x N ], where x i Represents the i-th observed variable. In order to make the indicator meet the assumptions of the algorithm, the data needs to be standardized. The specific steps are as follows:

[0081] For each variable x i Transformed to zero mean, that is, for each x i Subtract the mean:

[0082]

[0083] Make sure the mean of the data is 0.

[0084] Next, we use higher-order cumulants to calculate the causal difference τ between each pair of variables. ij . Higher-order cumulants are important tools for measuring linear correlations between random variables. Here, fourth-order cumulants and mixed cumulants are used to capture causal relationships in the data:

[0085] τ ij =|C4(x i )C 1,3 (x i ,x j )-C 2,2 (x i ,x j )C 3,1 (x i ,x j )|

[0086] Among them, the fourth-order accumulation C4(x i ) represents the measurement of a single variable x i The fourth-order statistical information of reflects its high-order structural characteristics. 1,3 (x i ,x j ) and C (2,2) (x i ,x j ) measure two variables x respectively i and x j The joint high-order features of is the accumulation amount, recorded as

[0087] Using the summation method, by calculating the sum of causal differences, we can quantify the causal influence of each performance indicator in the overall network, especially in the case of heteroscedastic noise. i Sum the causal differences with other indicators to get the causal difference of each indicator:

[0088] T τ (x i ; U) = ∑ {j∈U\i} τ ij .

[0089] Causal Difference and T τ (x i ; U) can help identify the causal effect of each indicator on other indicators under the influence of noise heteroskedasticity, and the T of the variable τ (x i The smaller the U), the less the indicator is affected by other indicators.

[0090] By identifying the index with the smallest causal difference, the root node in the causal relationship is determined. Based on the causal difference sum of each variable, the variable with the smallest causal difference sum is determined as the root node. The calculation formula is:

[0091]

[0092] Among them, x m It is the root node.

[0093] This application can effectively identify the root nodes that are least affected by other variables in the case of noise heteroskedasticity. For example, if the causal difference sum of bandwidth utilization is the smallest, this may indicate that it is the root cause of other performance indicators (such as latency or packet loss rate).

[0094] By calculating and iterating the residual matrix, the influence of exogenous variables is gradually removed, each root node is found level by level, and finally the causal order of wireless network performance indicators is determined.

[0095] At the first root node x m After being identified, it is known that this variable is the least affected by other variables in the causal structure, and is therefore considered to be the starting point of the causal relationship. At this point, it is necessary to separate the influence of this root node from the data and obtain the residual matrix of the remaining variables.

[0096] Determine the root node x m After that, we need to remove the influence of the root node on other variables from the data matrix X. Assume that the current data matrix X contains N variables, x m is the first root node identified, then the calculation steps of the residual matrix R are as follows:

[0097] For each non-root node x i (i≠m) is regressed, assuming it is the root node x m A linear function of:

[0098] x i =b mi x m -∈ i

[0099] Here mi is the regression coefficient, ∈ i is the residual, that is, after removing the root node x m After the influence of i The remaining part.

[0100] Through the regression equation, we get the residual ∈ i , these residuals form the new residual matrix R:

[0101] R=[∈1,∈2,…,∈ m-1 ,∈O+1 ,…,∈ N ]

[0102] In this new matrix R, each column represents a variable in the original data matrix after removing the root node x. m The residual after the influence of .

[0103] In order to continue to maintain the consistency of the data, the residual matrix R is standardized to zero mean

[0104] After obtaining the residual matrix R, repeat steps 2 to 4 to identify the next root node. The specific process is as follows: Use the residual matrix R to recalculate the causal difference τ of each variable ij Its causal difference and T τ (x i ; U); Find the variable x with the smallest causal difference m′ , and regard it as the new root node; remove the root node x again through linear regression m′ The influence of each root node on other variables is obtained by calculating the updated residual matrix R′; repeat the above steps to gradually remove the influence of each root node on other variables until all variables are identified as root nodes in turn.

[0105] Through multiple iterations, the causal order of all wireless network performance indicators can be finally determined. This method is not affected by noise heteroscedasticity on the causal structure, making the identified causal relationship more accurate and reliable.

[0106] In the method based on high-order accumulation, by calculating the causal difference between each variable and other variables, the noise heteroskedasticity problem that may exist in the data can be effectively handled. Specifically, the causal difference between the exogenous variable (root node) and the endogenous variable (leaf node) has causal order consistency, that is, the variable with the smaller causal difference is more likely to be the root node in the causal structure.

[0107] This consistency provides an important basis for the identification of causal structures, so that even when there is noise heteroscedasticity in the performance indicator data, the method of the present application can still accurately identify the causal relationship. Traditional causal discovery methods often do not have this consistency, which may lead to deviations or errors in the identification of causal structures when faced with noise heteroscedasticity.

[0108] Therefore, the method based on high-order accumulation can still maintain a high recognition effect under the condition of noise heteroscedasticity, effectively reveal the true causal relationship between wireless network performance indicators, and provide a more reliable basis for network performance optimization.

[0109] Through this iterative processing method of the residual matrix, the key causal relationships in wireless network performance indicators can be effectively identified, even in the presence of significant noise heteroskedasticity in the data, which provides a solid foundation for subsequent network performance optimization.

[0110] Through the identified root nodes and their causal order, the data causal structure of the entire wireless network performance indicators is gradually restored, thereby clarifying the causal relationship between the various performance indicators;

[0111] After multiple rounds of residual matrix calculation and root node identification, the causal order of wireless network performance indicators has been obtained. This order reflects the hierarchical relationship of various performance indicators in the causal structure, that is, the variable that is first identified as the root node is at the starting point of the causal chain, and the variables identified subsequently are affected by the previous variables in turn.

[0112] In order to systematically recover the causal structure, a lower triangular matrix B is constructed using the causal order. Each element of this matrix B ij Represents variable x i x j The matrix B is an N×N matrix, where N is the number of performance indicators. The elements in the matrix B ij The definition is as follows: If i>j, then B ij Indicates x in the causal order i x j The direct causal influence of ij =0, indicating x i x j There is no direct causal effect, because in the causal order, x i Cannot affect x that comes after it j .

[0113] For example, if the causal order is x1→x2→x3, then the structure of the lower triangular matrix is:

[0114]

[0115] In order to determine the specific values ​​of the non-zero elements in the matrix B (i.e., the causal connection coefficient), it is necessary to perform regression analysis on each performance indicator. j (j>1), regression analysis is performed to estimate the influence of its causal parent node. Assume x j The causal parent node is {x1,x2,…,x j-1}, the regression model can be expressed as:

[0116]

[0117] b ji is the regression coefficient to be estimated, indicating xi x j After completing the construction of the lower triangular matrix B and determining the connection coefficient, the causal structure between the wireless network performance indicators can be restored through the matrix.

[0118] After determining the final causal structure, the causal relationship in the wireless network performance indicator data can be obtained based on the causal graph. The application of causal relationships is of great significance in wireless network performance optimization, troubleshooting, prediction and control. Through detailed causal structure analysis, we can fundamentally understand the mutual influence between network performance indicators, thus providing strong support for practical applications.

[0119] This application can use high-order accumulation to establish a two-variable causal relationship discovery method and a multivariate causal relationship discovery algorithm to aggregate the two-variable causal relationship when the noise of wireless network performance indicator data is heteroscedastic, and discover the causal structure in the observation data with heteroscedastic noise. At the same time, because the causal difference is based on the accumulation, and the finite-order accumulation can be calculated by the finite-order moment, the complexity is reduced, but it still has a good recognition effect, and can more accurately and quickly discover the causal mechanism of the observation variable with heteroscedastic noise, providing strong help for the scientific discovery of subsequent wireless network performance optimization problems and wireless network maintenance.

[0120] Example 3

[0121] This embodiment proposes a causal discovery system for wireless network performance indicators, which is used to implement a causal discovery method for wireless network performance indicators proposed in Embodiment 1.

[0122] Figure 3 This is an overall framework diagram of the causal discovery system for wireless network performance indicators of this embodiment.

[0123] The causal discovery system for wireless network performance indicators includes:

[0124] A collection module, used to collect performance indicator data of several observed variables of the wireless network to form wireless network performance indicator data;

[0125] A causal difference calculation module, used to calculate the causal difference between the observed variables of the wireless network performance indicator data using high-order accumulation;

[0126] A root node search module, used to search for a root node of a causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference;

[0127] The causal structure recovery module is used to regard the order in which the root nodes are found as the causal order, and to recover the causal structure of the wireless network performance indicator data according to the causal order.

[0128] This embodiment proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the causal discovery method for wireless network performance indicators described in Example 1.

[0129] It can be understood that the causal discovery system and computer device for wireless network performance indicators of this embodiment improves the method of Example 1, and the options in the above-mentioned Example 1 are also applicable to this embodiment, so they will not be repeated here.

[0130] The same or similar reference numerals correspond to the same or similar components;

[0131] The terms used to describe the positional relationship in the drawings are only used for illustrative purposes and should not be construed as limiting the present embodiment;

[0132] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A causal discovery method for wireless network performance indicators, characterized in that: The following steps are involved: Collecting performance indicator data of several observation variables of the wireless network to form wireless network performance indicator data; Calculating causal differences between observed variables of the wireless network performance indicator data using high-order accumulation quantities; Searching for a root node of a causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference; The order in which the root nodes are searched is regarded as the causal order, and the causal structure of the wireless network performance indicator data is restored according to the causal order.

2. The causal discovery method for wireless network performance indicators according to claim 1, characterized in that: The expression of the wireless network performance indicator data includes: X=[x1,x2,…,x i ,…,x N ] In the formula, x i represents the performance indicator data of the i-th observed variable of the wireless network, and N represents the total number of observed variables; Before using the high-order accumulation to calculate the causal difference between the observed variables of the wireless network performance indicator data, the wireless network performance indicator data is standardized, and the expression of the standardized processing includes: In the formula, represents the normalized x i ,mean(x i ) represents x i The mean of .

3. The causal discovery method for wireless network performance indicators according to claim 1, characterized in that: The expression for calculating the causal difference between the observed variables of the wireless network performance indicator data using the high-order accumulation includes: τ ij =|C4(x i )C 1,3 (x i ,x j )-C 2,2 (x i ,x j )C 3,1 (x i ,x j )| C4(x i )=Cum(x i ,x i ,x i ,x i ) C 1,3 (x i ,x j )=Cum(x i ,x j ,x j ,x j ) C 2,2 (x i ,x j )=Cum(x i ,x i ,x j ,x j ) C 3,1 (x i ,x j )=Cum(x i ,x i ,x i ,x j ) In the formula, τ ij Represents the i-th observed variable x of the wireless network performance indicator data i and the jth observed variable x j ; Cum(·) represents the high-order cumulant function.

4. The causal discovery method for wireless network performance indicators according to claim 1, characterized in that: The step of searching for the root node of the causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference comprises: S1: For each observed variable in the wireless network performance indicator data, calculate the sum of the causal differences between it and other observed variables to obtain the sum of the causal differences corresponding to each observed variable; S2: The causal difference and the smallest observed variable are regarded as the root node x of the causal structure corresponding to the wireless network performance indicator data m ; S3: Remove the root node x from the wireless network performance indicator data. m The impact on other observed variables, and the residual matrix R is obtained; S4: Taking each observed variable in the residual matrix R as the operation object, iteratively execute steps S1 to S3, search for the root node level by level, and record the order in which the root nodes are searched, until all observed variables are identified as root nodes, stop the iteration, and obtain the causal order of the wireless network performance indicator data.

5. The causal discovery method for wireless network performance indicators according to claim 4, characterized in that: The expression of the causal difference sum includes: Where, T τ (x i ; U) represents x i The causal difference sum, U represents the subscript set of variables in the feature set X, and U\i represents removing the i-th variable x from the subscript set U. i .

6. The causal discovery method for wireless network performance indicators according to claim 4, characterized in that: The root node x m The expressions include: Where, T τ (x i ; U) represents x i The causal difference and.

7. The causal discovery method for wireless network performance indicators according to claim 1, characterized in that: The expression of the residual matrix R includes: R＝[∈1,∈2,…,∈ m-1 ,∈ m+1 ,…,∈ N ] x pi =b mi x m -∈ i In the formula, x m represents the root node; x pi represents the observed variable corresponding to the non-root node; b mi represents the regression coefficient; ∈ m-1 Represents the observed variable x m-1 The corresponding residual also indicates that the observed variable x is eliminated m-1 After the influence of m-1 The remaining part; ∈ i Represents x pi The corresponding residual; N represents the total number of observed variables.

8. The causal discovery method for wireless network performance indicators according to any one of claims 1 to 7, characterized in that: The step of restoring the causal structure of the wireless network performance indicator data according to the causal order comprises: Construct a lower triangular matrix B using the causal order, and let the element of the i-th row and j-th column of the lower triangular matrix B be B ij , if i≤j, then B ij =0, indicating x i x j There is no direct causal influence; Otherwise, B ij Represents the observed variable x in the causal sequence i For the observed variable x j The direct causal influence of j The causal parent node is {x1,x2,…,x j-1 }, then B ij The expressions include: B ij =b ji In the formula, ∈ j Represents x j The corresponding residual.

9. A causal discovery system for wireless network performance indicators, used to implement the causal discovery method for wireless network performance indicators as described in any one of claims 1 to 8, characterized in that: include: A collection module, used to collect performance indicator data of several observed variables of the wireless network to form wireless network performance indicator data; A causal difference calculation module, used to calculate the causal difference between the observed variables of the wireless network performance indicator data using high-order accumulation; A root node search module, used to search for a root node of a causal structure corresponding to the wireless network performance indicator data step by step based on the causal difference; The causal structure recovery module is used to regard the order in which the root nodes are found as the causal order, and to recover the causal structure of the wireless network performance indicator data according to the causal order.

10. A computer device comprising a memory and a processor, wherein the memory stores computer-readable instructions, characterized in that: When the computer-readable instructions are executed by the processor, the processor executes the steps of the causal discovery method for wireless network performance indicators according to any one of claims 1 to 8.

Images