B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold

Abstract

The invention relates to a B5G radio frequency front-end health parameter optimization method based on a bidirectional strategy and a self-adaptive threshold value, and belongs to the field of communication equipment health management. The method introduces associated information entropy to carry out global redundancy measurement aiming at the problems that high-dimensional monitoring parameters of a base station functional circuit are high in redundancy and key information is easy to submerge; carrying out parameter screening by fusing a reverse deletion strategy and a forward addition strategy; according to the method, core parameters are locked through reverse deletion, a preferred set is expanded through forward addition iteration, iteration termination is controlled through a self-adaptive threshold value, and the scale of the preferred set is determined autonomously; and finally, a parameter subset which is simple, low in redundancy and high in health characterization capability is automatically output, so that the efficiency and precision of subsequent health state evaluation are remarkably improved, and core technical support is provided for intelligent predictive operation and maintenance of the B5G base station.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of B5G base station health extraction characterization, and particularly relates to a B5G radio frequency front-end health parameter optimization method based on a bidirectional strategy and an adaptive threshold. BACKGROUND

[0002] With the evolution of the fifth generation mobile communication technology to B5G / 6G, the base station radio frequency front-end module (such as power amplifier, filter, low noise amplifier, etc.) as the key component of signal transceiver, its hardware structure is increasingly integrated and complex. In order to realize accurate health monitoring, multiple parameters including voltage, current, temperature, S parameter (scattering parameter), noise coefficient, stability coefficient, etc. need to be collected to form a high-dimensional health data space. These parameters together constitute the core information carrier reflecting the health status of the radio frequency front-end module, and their effective screening is the basis for implementing predictive maintenance and ensuring the stability of the communication link performance.

[0003] However, the existing parameter optimization method for radio frequency front-end modules has the following prominent limitations when facing the high-dimensional and strongly related monitoring parameters unique to radio frequency front-end modules:

[0004] First, the redundancy measurement perspective is limited, which is difficult to adapt to the complex correlation between radio frequency parameters. Traditional feature selection methods (such as MIFS, mRMR, etc.) mainly rely on the correlation between parameters and health status labels and the redundancy between parameters to select, which belongs to local pair-wise measurement. Multiple parameters of the radio frequency front-end module jointly represent the matching and transmission characteristics of the radio frequency link, and there is a complex physical coupling and nonlinear cooperative relationship between the parameters. The local view of the traditional method cannot quantify the overall information redundancy under the joint action of multiple parameters, which easily leads to the local optimization of the selected parameter subset, but fails to minimize the system-level redundancy;

[0005] Second, the parameter scale depends on manual preset, and lacks self-adaptation to circuit characteristics. Existing methods usually need to pre-set the number of parameters to be selected, which is mostly determined by expert experience. However, the health degradation mechanism and sensitive parameter types of different sub-circuits in the radio frequency front-end module are significantly different, and the number and combination of the required monitoring parameters are also different. The fixed-scale parameter selection strategy lacks flexibility and is difficult to adapt to the monitoring needs of different radio frequency circuits, reducing the engineering practicability of the method;

[0006] Third, the cost of full-parameter monitoring and processing is high, which is difficult to meet the real-time requirements. If the full parameter set of the radio frequency front-end module is directly collected, transmitted and modeled in real time, a large amount of hardware resources and network bandwidth will be consumed, and the data processing complexity is high, which is difficult to realize low-delay health status evaluation and fault warning on resource-limited base station equipment. SUMMARY

[0007] To address the aforementioned technical problems, this invention provides a method for optimizing B5G radio frequency (RF) front-end health parameters based on a bidirectional strategy and adaptive threshold. Applicable to the field of B5G base station RF front-end module health management, this optimization method quantifies parameter set redundancy from a system-wide perspective and adaptively determines the optimal selection scale. It also efficiently extracts parameters with low redundancy and high health sensitivity, supporting the construction of a refined and lightweight predictive maintenance system. Specifically, this invention introduces correlation information entropy as a global redundancy measurement tool for the high-dimensional monitoring parameter set of B5G base station RF front-ends. It innovatively integrates reverse deletion and forward addition strategies for parameter importance ranking and iterative screening, and introduces an adaptive threshold η to achieve intelligent and autonomous control of the optimal selection scale. This invention can automatically output a parameter optimization set that is streamlined, low in redundancy, and highly sensitive to health status, effectively improving the performance and efficiency of subsequent health feature extraction and status assessment models.

[0008] To achieve the above objectives, the present invention provides a method for optimizing B5G radio frequency front-end health parameters based on a bidirectional strategy and adaptive threshold, comprising the following steps:

[0009] S1: Construct the initial multidimensional health parameter set for the B5G base station radio frequency front-end module circuit;

[0010] S2: Calculate the correlation between each parameter and the preset health status category label based on mutual information theory, and standardize the data of the initial multidimensional health parameter set to calculate the correlation matrix of the standardized data;

[0011] S3: Calculate the association information entropy of the initial multidimensional health parameter set based on the eigenvalues ​​of the correlation matrix, as a global metric for measuring the overall redundancy of the parameter set;

[0012] S4: Execute the reverse deletion strategy: Iterate through each parameter in the initial parameter set, calculate the associated information entropy of the remaining parameter subset after removing the parameter, and calculate the resulting redundant information increment; select the parameter that maximizes the redundant information increment as the first core health-sensitive parameter, and add it to the preferred seed set;

[0013] S5: Execute the forward addition strategy: Based on the current preferred seed set, traverse the remaining parameters; for each remaining parameter, calculate the association information entropy of the new set formed after adding it to the current preferred seed set; select the parameter that maximizes the association information entropy value of the new set as the candidate addition parameter;

[0014] S6: Iteration termination judgment based on adaptive threshold η: Determine whether the entropy value of the new set associated with the candidate added parameter in S5 is less than the preset adaptive threshold η; if yes, terminate the iteration process; if no, formally add the candidate added parameter to the preferred seed set and return to S5 for the next round of iteration.

[0015] S7: When the iteration terminates, output the current preferred seed set as the final low-redundancy health-sensitive parameter preferred set.

[0016] Further, S1 includes the following steps:

[0017] S11: Based on the physical principles, working mechanisms, typical degradation mechanisms, degradation modes, and historical fault data of the RF front-end module circuit, select one or more candidate monitoring parameters from voltage, current, temperature, scattering parameters, noise figure, and stability coefficient that are sensitive to its health status evolution process, forming an initial multidimensional parameter set X = {x1, x2, ..., x...} n}, where n is the total number of parameters, and historical monitoring time series data for the entire state cycle are collected;

[0018] S12: Collect historical monitoring data of the RF front-end module circuit covering the entire lifecycle of health, degradation, danger, and failure, or the health degradation cycle, to form a dataset D∈R containing n variables and m time samples. m×n Each sample corresponds to a dataset with a health status label C, which includes healthy, degraded, dangerous, and faulty.

[0019] S13: Preprocess the data, including outlier removal, missing value imputation, and standardization or normalization of the time series of each parameter to eliminate the influence of dimensions and obtain a standardized data matrix.

[0020] Further, S2 includes the following steps:

[0021] S21: The mutual information I(X,C) between the parameter X and the health status label C, calculated based on information theory, is used to quantify the individual's ability to represent the health status. The formula for calculating the mutual information I(X,C) is as follows:

[0022] I(X,C)=H(X)+H(C)-H(X,C)

[0023] In the formula: H(X) is usually used to describe discrete random variables X = {x1, x2, ..., x...} n The entropy value of}; the formula for calculating H(X) is:

[0024]

[0025] In the formula: x i p(x) represents the possible values ​​of variable X. i ) is the probability density function;

[0026] When x is taken i and c i The joint probability is p(x) i ,c i When X and C are in the same condition, the formula for calculating the joint entropy of X and C is:

[0027]

[0028] S22: Suppose a multivariate system has n variables and collects m time samples. The multivariate time series matrix of the multivariate system at t (t=1,2,3,...,m) is P, then:

[0029] P=(y i (t)) 1≤t≤m,1≤i≤n ,P∈R m×n

[0030] Where: P is a real matrix; y i (t) represents the value of the i-th variable at time t;

[0031] After centering and standardizing matrix P, we obtain the multivariate time series matrix Q:

[0032] Q = (y i (t)) 1≤t≤m,1≤i≤n ,Q∈R m×n

[0033] The standardized data matrix is ​​centered, and the correlation matrix R between the parameters is calculated:

[0034]

[0035] Where R is an n×n symmetric positive semi-definite matrix with elements r ij Characterization parameter x i and x j The degree of linear correlation between them provides the basis for subsequent global redundancy measurement.

[0036] Further, S3 includes the following steps:

[0037] Calculate the global redundancy measure of the initial parameter set based on the correlation information entropy:

[0038] Using the correlation matrix R obtained from S2, calculate the association information entropy H of the entire initial parameter set R. R(X); The associated information entropy is an indicator that measures the overall information independence of a multivariable system, and the associated information entropy H R The formula for calculating (X) is:

[0039]

[0040] In the formula: Let be the i-th eigenvalue of the correlation matrix R calculated from standardized data; n is the total number of parameters; and H is the association information entropy. R The range of values ​​for (X) is [0,1], H R A larger (X) value indicates that the overall redundancy of the parameter set is lower and the information independence is stronger.

[0041] Simultaneously define a redundancy metric for this parameter set. The redundancy measure The calculation formula is:

[0042]

[0043] The closer it is to 0, the lower the overall redundancy of the parameter set.

[0044] Furthermore, step S4 executes a reverse deletion strategy to determine the first core health-sensitive parameter. This reverse deletion strategy aims to identify the parameter most closely associated with the overall health information of the parameter set and whose removal would cause the greatest disruption to the overall information structure, serving as the starting point. Specifically, it includes the following steps:

[0045] S41: For each parameter x in the initial parameter set X i Perform a virtual deletion operation to obtain subset X. miss(i) Calculate the association information entropy H of the remaining subset after removing it. R (X miss(i) );

[0046] S42: Calculate the removal parameter x i The resulting increase in redundant information ΔInfo i :

[0047] ΔInfo i =H R (X)-H R (X miss(i) );

[0048] S43: Because removing an important parameter usually leads to an increase in the redundancy of the remaining parameter set, i.e., H R (X) decreases, choose to make ΔInfo decrease i The parameter x with the largest value i*This means that removing it results in the greatest information loss, and its contribution to maintaining the overall information structure is the most crucial; that is, removing the parameter x... i* The parameter that causes the least damage to the overall information structure is considered a core parameter.

[0049] S44: x i* Add the preferred seed set S, i.e., S{x i*};

[0050] ΔInfo i The larger the value, the better the parameter x. i* The closer the correlation with the health status information represented by the entire parameter set, the greater the information loss caused by its absence, which manifests as increased redundancy.

[0051] Further, S5 includes the following steps:

[0052] S511: For the current preferred seed set S, iterate through all remaining candidate parameters x. j Based on the selected seed set S, this strategy iteratively selects parameters from the remaining parameter set U = XS that can bring the maximum effective information gain to the current preferred set, that is, maximize the increase of the associated information entropy and reduce redundancy.

[0053] S512: For each candidate parameter x j ∈U, calculate the new set s = {s∪{x} formed after adding it to S. j The entropy of the associated information H of}} R (S∪{x j});

[0054] S513: Select the option that enables H R (s∪{x j}) to reach the maximum candidate parameter x j* :

[0055]

[0056] S514: Record the maximum associated information entropy value H at this time. max_add (S)=H R (S∪{x j*}).

[0057] Further, S6 includes the following steps:

[0058] Each time the optimal candidate parameter x is found through S5... j* and its corresponding H max_add (S) is then used for judgment:

[0059] If H max_add If (S)>η, it indicates that parameter x has been added. j*Afterwards, if the overall independence level of the preferred set is still higher than or equal to the expected threshold, the effective information gain brought by this parameter is acceptable, and then x is... j* Formal inclusion in the preferred set S, i.e., S = S∪{x} j* Then return to S5 to continue the next iteration;

[0060] If H max_add If (S) < η, it indicates that even adding the current optimal candidate parameter cannot bring the association information entropy of the preferred set to the desired level of η. In this case, adding any remaining parameters will only introduce excessive redundant information, with a redundancy of η. Therefore, the algorithm automatically terminates the iteration;

[0061] Based on experimental analysis and verification, the recommended range for the threshold η is [0.60, 0.65]. Within this range, the algorithm can achieve the best balance between significantly reducing parameter redundancy (i.e., dimensionality reduction) and fully preserving key information about the health status.

[0062] To overcome the drawbacks of manually setting the number of parameters, this invention introduces an adaptive threshold η, which serves as an intelligent criterion for algorithm termination and includes the user's expectation of the minimum level of global information independence allowed by the final optimal parameter set.

[0063] Further, S7 includes the following steps:

[0064] When the algorithm terminates according to condition S6, the currently constructed preferred seed set S is the final result. The preferred seed set S has the following characteristics: its size is adaptively determined by the algorithm without manual intervention; low global redundancy among parameters within the set; and H... R (S) is close to η; the parameter has a strong characterization ability for the health status of the target functional circuit.

[0065] Innovation and beneficial effects of this invention:

[0066] 1. Global Redundancy Measurement Perspective: This innovative approach transfers the theory of correlated information entropy from the field of multi-sensor data fusion to the optimization problem of health parameters in B5G base station RF front-end modules. This theory, starting from the overall system, quantifies the redundancy information under the complex combined effects and synergistic correlations among multiple parameters, overcoming the fundamental limitation of traditional methods (such as MIFS, mRMR) which only perform local, pairwise measurements, thus providing a more comprehensive and accurate theoretical basis for optimization.

[0067] 2. Robust Ranking Mechanism Integrating Two Strategies: A robust ranking framework combining "reverse deletion initialization" and "forward addition expansion" was designed. Reverse deletion first anchors the most core parameters related to overall health information, ensuring the robustness of the optimization foundation; forward addition, on this basis, iteratively absorbs parameters that bring the maximum effective information increment to the current set. The synergy of the two strategies enables the ranking process to have both a global perspective and local optimization capabilities, effectively improving the quality of the optimization results.

[0068] 3. Intelligent Termination with Adaptive Scale Control: By introducing a threshold η as an adaptive criterion for algorithm termination, the optimal number of parameters can be determined automatically. The algorithm can automatically determine when to stop adding parameters based on the information structure of the data itself, achieving intelligent and objective determination of the optimal scale. This greatly enhances the method's universality and engineering practical value for different types of functional circuits (with varying parameter importance distributions).

[0069] 4. Outputting a high-performance parameter subset: The final optimized parameter subset S has significant advantages of "low redundancy, high correlation, and simplified scale." This directly brings three benefits: effectively reducing the data dimensionality and computational overhead of subsequent health feature extraction and state modeling; reducing redundant noise interference and improving the accuracy and generalization ability of the health assessment model; and providing key front-end data processing technology support for the low-cost, high-precision intelligent predictive operation and maintenance of B5G base station radio frequency front-end modules.

[0070] Table 1 shows a comparison of the expected effects of the method described in this invention with traditional parameter optimization methods:

[0071] Table 1:

[0072]

[0073] Attached image description:

[0074] Figure 1 This is an overall flowchart of a preferred scheme of the method for optimizing B5G radio frequency front-end health parameters based on bidirectional strategy and adaptive threshold according to the present invention;

[0075] Figure 2 The flowchart shows the sequence and loop of the preferred scheme of the method for optimizing B5G radio frequency front-end health parameters based on bidirectional strategy and adaptive threshold according to the present invention.

[0076] Figure 3 This is a heatmap showing the correlation of B5G base station radio frequency front-end module parameters in a preferred scheme of the B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention.

[0077] Figure 4This is a parameter degradation curve of the B5G base station radio frequency front-end module in a preferred scheme of the B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention.

[0078] Figure 5 This is a ranking diagram of the importance of B5G base station radio frequency front-end module parameters in a preferred scheme of the B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention.

[0079] Figure 6 This is a schematic diagram of the correlation information entropy calculation principle based on feature values ​​in the preferred scheme of the B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention.

[0080] Figure 7 The flowchart shows the reverse deletion sequence and loop in the preferred scheme of the B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention.

[0081] Figure 8 This is a flowchart illustrating the order and loop of forward addition in the preferred scheme of the preferred method for optimizing B5G radio frequency front-end health parameters based on bidirectional strategy and adaptive threshold of the present invention.

[0082] Figure 9 A radar chart showing the comprehensive performance comparison between the preferred scheme of the B5G RF front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention and existing schemes;

[0083] Figure 10 This is a performance comparison diagram between the preferred scheme of the B5G RF front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention and the existing scheme;

[0084] Figure 11 This is a comparison diagram of the preferred parameter set in the preferred scheme of the B5G radio frequency front-end health parameter optimization method based on bidirectional strategy and adaptive threshold of the present invention with the traditional MIFS method and the existing full parameter set in terms of health degradation characterization. Detailed implementation method:

[0085] To solve the above-mentioned problems in the prior art, the following is combined with the appendix of the present invention. Figures 1-11 The preferred embodiments of the present invention will be described in detail below.

[0086] This invention provides a method for optimizing B5G radio frequency front-end health parameters based on a bidirectional strategy and adaptive threshold, such as... Figure 1 and Figure 2 As shown, the specific implementation steps are as follows:

[0087] Step 1: Constructing the initial parameter set and data preparation, specifically including the following steps:

[0088] Step 1.1: Determine the target functional circuit and candidate parameters: This embodiment takes the radio frequency front-end power amplifier (PA) module in the B5G base station AAU as the research object. The degradation curves of each parameter are as follows: Figure 4 As shown, the performance degradation of the PA module directly affects signal transmission quality and efficiency. Based on RF circuit theory and engineering experience, the following six key electrical parameters that are sensitive to the PA's health status are selected to form the initial parameter set X:

[0089] X={x1:S 11 ,x2:S 12 ,x3:S 21 ,x4:S 22 ,x5:NF,x6:Stab}

[0090] Wherein: S 11 S 12 S 21 S 22 The scattering parameter (S) coefficients represent the input reflection coefficient, reverse transmission coefficient, forward transmission gain, and output reflection coefficient, respectively; NF is the noise figure; Stab is the stability coefficient; the correlation between the PA module parameters is referenced. Figure 3 ;

[0091] Step 1.2: Data health status classification preprocessing: Based on the degree of decay of key performance indicators of the module, including output power and efficiency, each sample is labeled with four health status labels C: healthy (H), degraded (D), dangerous (R), and faulty (F);

[0092] Step 1.3: Data preprocessing, including:

[0093] Outlier handling: The 3σ principle is used to remove outlier data points caused by transient interference during measurement;

[0094] Missing value imputation: Time series linear interpolation is used to fill in the small amount of missing data caused by occasional sensor failures;

[0095] Data standardization: To eliminate the influence of differences in the dimensions and numerical ranges of various parameters, Z-score standardization is performed on the time series data of each parameter, making its mean 0 and standard deviation 1. The processed data constitute a standardized data matrix D. std ∈R 500×6 ;

[0096] Step 2: Calculate the mutual information and correlation matrix, which includes the following steps:

[0097] Step 2.1: Calculate parameter-health status mutual information:

[0098] First, the continuous parameter values ​​are discretized and binned, and the health status label is treated as a discrete variable. Then, each parameter x is calculated. i Mutual information I(x) with health status C i ,C), with parameter S 12 For example:

[0099] I(S 12 ,C)=H(S 12 )+H(C)-H(S 12 C)

[0100] in: In the formula For S 12 The probability that the value falls into the b-th bin; H(C) and joint entropy H(S) 12 C) Similar calculations; the calculated mutual information values ​​reflect the amount of health status information carried by each parameter individually;

[0101] Step 2.2: Construct the correlation matrix between parameters: For the standardized data matrix D std After centering, we obtain matrix Q. We then calculate the 6×6 Pearson correlation matrix R (m=500):

[0102]

[0103] Where: r ij Indicates parameter x i and x j The linear correlation coefficients between them; this matrix characterizes the pairwise linear relationships between all parameters;

[0104] Step 3: Calculate the global correlation information entropy, referring to... Figure 6 The principle diagram for calculating the entropy of related information includes the following steps:

[0105] Step 3.1: Calculate the eigenvalues ​​of the correlation matrix: Solve the characteristic equation R-λI|=0 of the correlation matrix R to obtain 6 eigenvalues. Since R is a real symmetric matrix, all of its eigenvalues ​​are non-negative.

[0106] Step 3.2: Calculate the entropy of the associated information of the initial parameter set:

[0107]

[0108] In this embodiment, H is calculated R If (X) = 0.382, then the redundancy of the initial parameter set is... The high redundancy verifies the necessity of dimensionality reduction for all six parameters;

[0109] Step 4: Reverse deletion to determine core parameters, such as Figure 7As shown, the specific steps include:

[0110] Step 4.1: Calculate the entropy of the association information after removing each parameter: Remove each parameter x from the initial set X. i After removal, we get subset X miss(i) For example, removing S 12 The subsequent subset is Recalculate the correlation matrix and eigenvalues ​​of the five parameter subsets in the formula, and then calculate their association information entropy. Similar calculations and H R (X miss(NF) );

[0111] Step 4.2: Calculate the redundancy information increment and select core parameters:

[0112] Step 4.2.1: Calculate the increase in redundant information caused by removing each parameter:

[0113] ΔInfo i =H R (X)-H R (X miss(i) )

[0114] refer to Figure 5 The results of the parameter importance ranking calculation are shown in Table 2:

[0115] Table 2:

[0116]

[0117]

[0118] Step 4.2.2: Select ΔInfo i Maximum (i.e., H) R (X miss(i) The minimum parameters are shown in Table 2. Remove S. 12 hour Maximum means S 12 The absence of [something] leads to the greatest increase in redundancy of the remaining parameter set and the lowest decrease in the entropy of the associated information, indicating that S [something]. 12 It is the most crucial parameter for maintaining the overall parameter set information structure, therefore S 12 Selected as the first health-sensitive parameter, the initialization seed set S = {S 12};

[0119] Steps 5 and 6: A loop of forward addition and adaptive threshold judgment, as follows... Figure 8 As shown:

[0120] The adaptive threshold η is set to 0.62, which is within the recommended range of [0.60, 0.65].

[0121] First iteration:

[0122] 1. Candidate parameters: U = XS = {S} 11 ,S 21 ,S 22 ,NF,Stab};

[0123] 2. Calculation and Selection:

[0124] Calculate the association information entropy of the new set after adding each candidate parameter to S:

[0125] H R ({S 12 ,S 11})=0.580

[0126] H R ({S 12 ,S 21})=0.595

[0127] H R ({S 12 ,S 22})=0.592

[0128] H R ({S 12 ,NF})=0.669

[0129] H R ({S 12 ,Stab})=0.520

[0130] Wherein: H R ({S 12 The maximum value is η(0.62) = 0.669, and 0.669 ≥ η(0.62).

[0131] 3. Decision: If the threshold condition is met, add NF to the preferred set and update S = {S} 12 ,NF};

[0132] Second iteration:

[0133] 1. Candidate parameters: U = XS = {S} 11 ,S 21 ,S 22 ,Stab};

[0134] 2. Calculation and Selection:

[0135] Calculate separately:

[0136] HR ({S 12 ,NF,S 11})=0.600

[0137] H R ({S 12 ,NF,S 21})=0.610

[0138] H R ({S 12 ,NF,S 22})=0.592

[0139] H R ({S 12 ,NF,Stab})=0.543

[0140] Where: the maximum value H max_add (S)=H R ({S 12 ,NF,S 21})=0.610;

[0141] 3. Decision: At this point, H max_add (S)=0.610<η(0.62); According to the algorithm rules, even if the best candidate parameter S is added... 21 It is also impossible to make the association information entropy of the preferred set reach the expected threshold η. This indicates that the effective information gain brought by adding any parameters is not enough to offset the redundancy it introduces. Therefore, the algorithm terminates automatically without adding any new parameters.

[0142] Step 7: Output the final optimal set:

[0143] Algorithm performance comparison Figure 9 , Figure 10 and Figure 11 As shown;

[0144] After the algorithm terminates, the final optimal set of health-sensitive parameters is S = {S 12 The NF method contains only 2 parameters, which significantly reduces dimensionality compared to the initial 6 parameters, by 66.7%.

[0145] Verification and analysis of the effects of this preferred embodiment:

[0146] To quantitatively evaluate the effectiveness of this invention, the following comparative experiments were conducted:

[0147] 1. Comparison method:

[0148] Full parameter set: using all 6 parameters {S} 11 ,S 12 ,S 21 ,S 22,NF,Stab}.

[0149] Traditional MIFS method: This method applies traditional feature selection based on mutual information, pre-selecting three parameters, and its output is {S}. 21 ,S 11 ,NF}.

[0150] 2. Evaluation Indicators:

[0151] Redundancy metric: Calculate the entropy H of the associated information of each parameter subset. R and redundancy

[0152] Health status classification accuracy: Using the same random forest classifier and the same hyperparameters, the features extracted from the above three parameter sets (using the health degradation coefficient of the parameters) are used as input to classify the four states of health, degradation, danger and failure. Five-fold cross-validation is used, and the average classification accuracy is used as the evaluation index.

[0153] 3. The experimental results are shown in Table 3:

[0154] Table 3:

[0155]

[0156] 4. Results Analysis:

[0157] Redundancy: The preferred set of this invention has the highest association information entropy (0.669) and the lowest redundancy (0.331), which is 46.4% lower than the full parameter set and 22.1% lower than the traditional MIFS method, proving that this invention can more effectively remove redundant information.

[0158] Characterization capability: The health status classification accuracy achieved by this invention using only two parameters (91.8%) is almost on par with that using the full parameter set of all six parameters (92.4%), and significantly higher than that of the traditional MIFS method (87.6%), indicating that the optimal two parameters {S} 12 ,NF} can fully and efficiently characterize the health degradation process of PA modules with very little loss of useful information.

[0159] In summary, this embodiment demonstrates in detail the complete application process and outstanding effects of the present invention on the B5G base station radio frequency front-end module. The present invention achieves intelligent and precise optimization of health-sensitive parameters through rigorous theoretical foundation and adaptive algorithm design.

[0160] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for B5G RF front-end health parameter optimization based on bidirectional strategy and adaptive threshold, characterized in that, Comprising the following steps: S1: Constructing an initial multi-dimensional health parameter set of a B5G base station radio frequency front-end module circuit; S2: Calculating the correlation between each parameter and the preset health state category label based on the mutual information theory, and performing standardization processing on the data of the initial multi-dimensional health parameter set, and calculating the correlation matrix of the standardized data; S3: According to the eigenvalue of the correlation matrix, the correlation information entropy of the initial multi-dimensional health parameter set is calculated as a global measurement index for measuring the overall redundancy of the parameter set; S4: Perform a reverse deletion strategy: traverse each parameter in the initial parameter set, calculate the correlation information entropy of the remaining parameter subset after removing the parameter, and calculate the redundancy information increment generated thereby; select the parameter that maximizes the redundancy information increment as the first core health sensitive parameter, and add it to the preferred seed set; S5: Perform a forward addition strategy: based on the current preferred seed set, traverse the remaining parameters; for each remaining parameter, calculate the correlation information entropy of the new set formed after adding it to the current preferred seed set; select the parameter that can maximize the correlation information entropy value of the new set as the candidate addition parameter; S6: Iteration termination judgment based on an adaptive threshold η: judge whether the correlation information entropy value of the new set corresponding to the candidate addition parameter in S5 is less than the preset adaptive threshold η; if yes, terminate the iteration process; If not, add the candidate addition parameter to the preferred seed set, and return to S5 for the next iteration; S7: When the iteration is terminated, output the current preferred seed set as the final low-redundancy health sensitive parameter preferred set.

2. The B5G RF frontend health parameter preferred method based on bidirectional policy and adaptive threshold of claim 1, wherein, The S1 comprises the following steps: S11: According to the physical principle, working mechanism, typical degradation mechanism, degradation mode of the radio frequency front-end module circuit and historical failure data, one or more candidate monitoring parameters of voltage, current, temperature, scattering parameter, noise coefficient and stability coefficient which are sensitive to the health state evolution process of the radio frequency front-end module circuit are screened out, and an initial multi-dimensional parameter set X={x1, x2,..., xN} is formed, wherein N is the total number of parameters, and historical monitoring time series data of the whole state cycle is collected; n} S12: Collecting historical monitoring data of the RF front-end module circuit covering the complete life cycle or health degradation cycle of health, degradation, danger and failure, forming a data set D e R containing n variables, m time samples m×n , each sample corresponding to a health status label C, the health status label C including health, degradation, danger and failure; S13: Preprocessing the data, the preprocessing including outlier rejection, missing value interpolation, and standardization processing or normalization processing on each parameter time series to eliminate the dimension effect, to obtain a standardized data matrix.

3. The B5G RF frontend health parameter preferred method based on bidirectional policy and adaptive threshold of claim 1, wherein, The S2 comprises the following steps: S21: Calculate the mutual information I(X,C) between the parameter X and the health state label C based on information theory, and the calculation formula of I(X,C) is: I(X,C) = H(X) + H(C) - H(X,C) In the formula: H(X) is usually used to describe discrete random variables X = {x1, x2, ..., x...} n The entropy value of}; The calculation formula of H(X) is: where: x i are possible values of the variable X; p(x i ) is the probability function; When taking the joint probability of x i and c i as p(x i , c i ), the joint entropy of X and C is calculated as follows: S22: Assuming that a multivariate system has n variables, m time samples are collected, and the multivariate time series matrix of the multivariate system at t (t = 1, 2, 3, …, m) is P, then: P = (y i (t)) 1≤t≤m,1≤i≤n ,P∈R m×n where: P is a real matrix; y i (t) denotes the value of the i-th variable at time t; After centering and standardizing the matrix P, a multivariate time series matrix Q is obtained: Q = (y i (t)) 1≤t≤m,1≤i≤n Q∈R m× n Centering the standardized data matrix, the correlation matrix R between parameters is calculated: where: R is an n x n symmetric positive semi-definite matrix whose elements r ij characterization parameter x i and x j between the linear correlation degree, this matrix provides the basis for subsequent global redundancy measurement.

4. The B5G RF frontend health parameter preferred method based on bidirectional policy and adaptive threshold of claim 1, wherein, The S3 comprises the following steps: Based on the correlation information entropy, the global redundancy measurement of the initial parameter set is calculated: Using the correlation matrix R obtained by S2, the correlation information entropy H of the whole initial parameter set R is calculated R (X); the correlation information entropy is an index for measuring the overall information independence of a multivariate system, and the calculation formula of the correlation information entropy H R (X) is: In the formula: is the ith eigenvalue of the correlation matrix R calculated for the standardized data; n is the total number of parameters; the associated information entropy H R The value range of (X) is [0, 1], H R The greater the value of (X) indicates that the overall redundancy of the parameter set is lower, and the information independence is stronger. to define a redundancy measure of the parameter set the redundancy measure The formula is: The closer to 0, the lower the overall redundancy of the parameter set.

5. The B5G RF frontend health parameter preferred method based on bidirectional policy and adaptive threshold of claim 1, wherein, The S4 performs a reverse deletion strategy to determine the first core health sensitive parameter, and the reverse deletion strategy aims to find the parameter that is most closely related to the health information of the entire parameter set and causes the most damage to the overall information structure after being removed as a starting point, and specifically comprises the following steps: S41: for each parameter x in the initial parameter set X i , a virtual deletion operation is performed to obtain a subset X miss(i) , the associated information entropy H R (X miss(i) ) of the remaining subset after removing x is calculated; S42: Calculate removal parameter x i Caused redundancy information increment ΔInfo i : Delta Info i = H R (X)-H R (X miss(i) ); S43: Since removing an important parameter usually results in the redundancy of the remaining parameter set rising, i.e. H R (X) decreases, select the parameter x i with the largest value of ΔInfo i* , meaning that removing it results in the largest loss of information and its contribution to maintaining the overall information structure is the most core, i.e. removing the parameter x i* destroys the overall information structure the least, indicating that it is a core parameter; S44: select x i* Add the preferred seed set S, i.e. S{x i*} ΔInfo i The larger, the more the parameter x i* The closer the association with the health status information represented by the entire parameter set, the greater the information loss resulting from its absence, which manifests as an increase in redundancy.

6. The B5G RF frontend health parameter preferred method based on bidirectional policy and adaptive threshold of claim 1, wherein, The S5 comprises the following steps: S511: For the current preferred seed set S, traverse all the remaining candidate parameters x j On the basis of the selected seed set S, the strategy iteratively selects the parameters from the remaining parameter set U=X-S that can bring the maximum effective information gain to the current preferred set, that is, the maximum degree of improvement of the relevant information entropy and the reduction of redundancy. S512: For each candidate parameter x j ∈ U, calculate the association information entropy H j}} of the new set S = {S ∪ {x R , x j}} after adding x S513: select a candidate parameter x that maximizes H R (S∪{x j}) over all candidate parameters x j* : S514: record the maximum association information entropy value H at this time max_add (S) = H R (S U {x j*}).

7. The bidirectional policy and adaptive threshold based B5G RF front-end health parameter preferred method of claim 1, wherein, The S6 comprises the following steps: At each pass the S5 finds the best candidate parameter x j* and its corresponding H max_add (S) a decision is made: If H max_add (S) > η, then the parameter x j* After, if the overall independence level of the preferred set is still higher or equal to the desired threshold, the effective information gain brought by the parameter is acceptable, and x j* is formally added to the preferred set S, i.e. S = S U {x j*}, and the process returns to S5 to continue the next iteration. If H max_add (S) < η, then it indicates that even adding the current optimal candidate parameter cannot make the association information entropy of the preferred set reach the desired η level, at which time adding any remaining parameters will only introduce excessive redundant information, with a redundancy degree of The algorithm thus automatically terminates iteration; According to the experimental analysis and verification, the recommended value range of the threshold η is [0.60, 0.65], and in the value range, the algorithm can achieve the best balance between significantly reducing parameter redundancy, that is, realizing dimension reduction, and sufficiently retaining key information of the health state.

8. The B5G RF frontend health parameter preferred method based on bidirectional policy and adaptive threshold of claim 1, wherein, The adaptive threshold η in the S6 is a preset threshold parameter, which is used to control the termination timing of the optimization process.

9. The bidirectional policy and adaptive threshold based B5G RF front-end health parameter preferred method of claim 1, wherein, The S7 includes the following steps: When the algorithm terminates according to the condition of S6, the preferred seed set S currently constructed is the final result, which has the following characteristics: the size is adaptively determined by the algorithm, without human intervention; the global redundancy among the parameters in the set is low, H R (S) Close to η; the parameters have strong characterization ability for the health state of the target functional circuit.

Images