CNN-Kriging model-based index system dynamic construction method and device

Abstract

The invention relates to an index system dynamic construction method and device based on a CNN-Kriging model. The method is applied to a center node of a distributed intelligent spectrum access system comprising the center node and a plurality of distributed edge nodes. Comprising the following steps: preprocessing a data sample reported by a distributed edge node, dividing a preprocessing result into a training set and a test set, and training a CNN-Kriging model; a synthetic data set is generated according to the trained CNN-Kriging model; and performing statistical analysis and index optimization by adopting a global sensitivity analysis method according to the synthetic data set, updating a data sample according to an optimized index system, and updating the index system to obtain an optimized index system oriented to a spectrum access decision. According to the method, the problems of low spectrum efficiency, poor reliability and incapability of self-adaption to a dynamic environment caused by a static and uninterpretable interference management strategy under dense deployment of massive Internet of Things equipment are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and relates to a CNN-Kriging model-based index system dynamic construction method and device. BACKGROUND

[0002] With the rapid development of Internet of Things (IoT) technology, a large number of IoT devices (such as sensors, actuators, tags, etc.) are densely deployed in smart factories, smart agriculture, smart cities and other scenarios. These devices usually work in unlicensed or shared spectrum (such as ISM band, unlicensed 5GHz band), and their large-scale random access behavior can cause complex co-frequency and adjacent frequency interference, which seriously restricts the overall network capacity and transmission reliability.

[0003] Existing spectrum access and interference management schemes mainly fall into three categories: the first category is a distributed contention protocol based on carrier sense multiple access (CSMA / CA), which is essentially "listen before talk". In dense deployment, the hidden terminal and collision problems are prominent, and the efficiency is low; the second category is dynamic spectrum access (DSA) based on fixed rules, which relies on a preset interference map or fixed strategy library, lacks adaptive ability to dynamic and complex environments, and the update lags behind; the third category is a centralized optimization method, which needs to gather all information to the center node for global optimization, has large decision delay, and requires extremely high computing and communication capabilities of the center node, which is difficult to expand.

[0004] In recent years, machine learning technology, especially deep learning, has been introduced to improve the intelligence of spectrum sensing and decision-making. For example, convolutional neural networks (CNN) are used to identify spectrum holes, or reinforcement learning (RL) is used to learn access strategies. However, existing methods have the following limitations: 1) the model is mostly a "black box", the decision-making process is not interpretable, and it is difficult to apply and trust in industrial scenarios with high reliability requirements; 2) feature engineering relies on expert experience, and the constructed index system is static and fixed, and cannot adapt to the dynamic changes of interference causes in different regions and different business scenarios; 3) there is a lack of quantification of the uncertainty of the prediction results, resulting in uncontrollable decision-making risks. SUMMARY

[0005] In view of the problems existing in the above-mentioned traditional method, the application proposes a CNN-Kriging model-based index system dynamic construction method and device.

[0006] To achieve the above purpose, the embodiments of the application adopt the following technical solutions: On the one hand, a CNN-Kriging model-based index system dynamic construction method is provided, which is applied to a center node in a distributed intelligent spectrum access system comprising a center node and a plurality of distributed edge nodes, comprising the steps of: Step 1: Obtain multiple groups of data samples reported by each distributed edge node; the data samples include spectrum environment data, device context data, and channel performance data.

[0007] Step 2: Preprocess the data samples, divide the preprocessing results into a training set and a test set; train the CNN-Kriging model using the training set, and then test it using the test set to obtain the trained CNN-Kriging model; the CNN-Kriging model includes: a CNN network and a Kriging module; the CNN network is used for feature extraction of training samples, and the Kriging module is used for inputting the extracted features to construct a Kriging proxy model.

[0008] Step 3: Generate a synthetic data set through active sampling based on the trained CNN-Kriging model.

[0009] Step 4: Perform statistical analysis and index optimization using a global sensitivity analysis method based on the synthetic data set to determine the optimized index system.

[0010] Step 5: Update the data samples according to the optimized index system.

[0011] Step 6: Repeat steps 2 to 5 until the preset condition is met to obtain the optimized index system for spectrum access decision.

[0012] On the other hand, an index system dynamic construction device based on a CNN-Kriging model is provided, comprising: A data sample acquisition module is configured to acquire multiple groups of data samples reported by each distributed edge node; the data samples include spectrum environment data, device context data, and channel performance data.

[0013] A CNN-Kriging model training module is configured to preprocess the data samples, divide the preprocessing results into a training set and a test set; train the CNN-Kriging model using the training set, and then test it using the test set to obtain the trained CNN-Kriging model; the CNN-Kriging model includes: a CNN network and a Kriging module; the CNN network is used for feature extraction of training samples, and the Kriging module is used for inputting the extracted features to construct a Kriging proxy model.

[0014] A synthetic data set generation module is configured to generate a synthetic data set through active sampling based on the trained CNN-Kriging model.

[0015] The index system optimization module is configured to perform statistical analysis and index optimization on the synthetic data set by using a global sensitivity analysis method, and determine an optimized index system; and update the data sample according to the optimized index system.

[0016] The optimization index system determination module for spectrum access decision is configured to repeatedly execute the CNN-Kriging model training module, the synthetic data set generation module, and the index system optimization module until a preset condition is met, and obtain an optimized index system for spectrum access decision.

[0017] One of the above technical solutions has the following advantages and beneficial effects: The above index system dynamic construction method and device based on the CNN-Kriging model are applied to a central node of a distributed intelligent spectrum access system including a central node and multiple distributed edge nodes; the method includes: preprocessing data samples reported by each distributed edge node, dividing the preprocessing results into a training set and a test set, and training a CNN-Kriging model; generating a synthetic data set according to the trained CNN-Kriging model; performing statistical analysis and index optimization on the synthetic data set by using a global sensitivity analysis method, updating the data sample according to the optimized index system, updating the index system, and obtaining an optimized index system for spectrum access decision. This method solves the problem of low spectrum efficiency, poor reliability, and inability to adapt to dynamic environments caused by static and uninterpretable interference management strategies under the dense deployment of massive Internet of Things devices. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A flowchart of an embodiment of the index system dynamic construction method based on the CNN-Kriging model; Figure 2 A schematic diagram of a CNN-Kriging model framework in an embodiment. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application.

[0022] It is to be understood that the phraseology or terminology employed herein, and not otherwise defined herein, is for the purpose of description only and is not intended to be limiting. Any use of section headings is intended to aid reading the application and is not to be interpreted as limiting. The details in the specification are for illustrative purposes only and are not intended to limit the scope of the application.

[0023] The embodiments of the present application will be described in detail with reference to the drawings, wherein the same reference numerals refer to the same elements throughout the several views. As this description travels from background, to detailed description of the drawings, it is understood that the application is not limited in scope to the examples that will be described.

[0024] In one embodiment, as shown in FIG. 1, a CNN-Kriging model-based index system dynamic construction method is provided, which is applied to a central node in a distributed intelligent spectrum access system including a central node and a plurality of distributed edge nodes, and can include the following processing steps 1 to 6: Figure 1 Step 1: Obtain a plurality of groups of data samples reported by each distributed edge node; the data samples include spectrum environment data, device context data, and channel performance data.

[0025] Specifically, the distributed intelligent spectrum access system includes a central node (such as a cloud server or a regional network controller) and a plurality of edge access gateways (i.e. distributed edge nodes) deployed in each region of the intelligent factory. Internet of Things devices access the network through their associated gateways.

[0026] The central node periodically collects data from each edge gateway, including: signal strength indication (RSSI), signal-to-noise ratio (SNR), and packet error rate on different channels perceived by each gateway; location information (such as obtained by triangulation or beacon), device type (such as sensor, AGV, mechanical arm), and service mode (periodic reporting, event triggering, streaming) of each access device; and actual throughput, latency, and other performance data of each channel.

[0027] ​Step 2: pre-process the data samples, divide the pre-processing results into a training set and a test set; train the CNN-Kriging model using the training set, and then test using the test set to obtain the trained CNN-Kriging model; the CNN-Kriging model includes: a CNN network and a Kriging module; the CNN network is used for feature extraction of the training samples, and the Kriging module is used for inputting the extracted features to construct a Kriging surrogate model.

[0028] Specifically, model training and updating. Based on the collected historical and real-time data, a CNN-Kriging model is trained. The input of the model is all feature variables in the data collected periodically by the central node from each edge gateway. The output of the model is the interference level prediction value (such as the normalized interference power) of the location on the channel and the Kriging variance of the prediction value (representing uncertainty). The CNN part is responsible for extracting deep spatial-spectral features from the combination of location and channel, and the Kriging part is responsible for interpolation prediction based on these features and spatial correlation.

[0029] A Kriging model is constructed which can accurately reflect the complex mapping relationship between the key input parameters and the output performance response of the system. Under the premise of ensuring sufficient prediction accuracy of the input and output behavior of the original model, the time required for a single evaluation is shortened from several hours or days to seconds or even milliseconds, thereby realizing large-scale and efficient calculation experiments. For example, support vector regression machine is often selected because of its good generalization ability in handling small samples and nonlinear problems; the GMDH method in self-organizing data mining algorithm can effectively handle high-dimensional complex systems and the model structure is transparent through automatic network structure growth and screening; Gaussian process regression can not only provide point estimates of prediction values, but also give uncertainty measures of prediction; and for problems with extremely complex input-output relationship and sufficient data, deep neural networks can construct a highly expressive surrogate model due to their strong nonlinear fitting ability and hierarchical feature extraction ability. Regardless of the specific technology used, the construction process usually follows the standard supervised learning process: first, run the full-order model to obtain a certain number of input-output sample data to form a training set, then train and cross-validate the selected surrogate model based on the training set, and finally ensure that the model reaches the preset prediction accuracy threshold on the independent test set, so that it can reliably replace the original complex model to perform fast performance deduction and prediction.

[0030] The CNN-Kriging model proposed by the method adopts end-to-end fusion of feature extraction and regression modeling, automatically extracts multi-scale local features of data through CNN, and realizes efficient regression using the spatial interpolation method of Kriging, avoiding the limitations of manual feature design.

[0031] The CNN-Kriging model framework is as shown in FIG. 1. Figure 2

[0032] Step 3: According to the trained CNN-Kriging model, synthetic data sets are generated by active sampling.

[0033] Step 4: According to the synthetic data sets, statistical analysis and index optimization are performed by using the global sensitivity analysis method to determine the optimized index system.

[0034] Specifically, using the trained CNN-Kriging model, within its input definition domain (the entire factory area, all channels), the Latin hypercube sampling method is used to generate an analysis data set containing tens of thousands of {input feature combinations -> model predicted output} samples. To improve efficiency, the sampling density can be increased in areas with high model prediction variance (i.e., areas with high model cognitive uncertainty).

[0035] The Sobol global sensitivity analysis method is used to analyze the generated data set. A set of candidate feature indicators is set, which may be derived from the original data, such as: same-channel device density, total power of adjacent-channel devices, distance from the nearest known strong interference source, proportion of same-type service devices, and regional average path loss. The first-order sensitivity index (independent contribution) and total sensitivity index (total contribution including interaction) of each candidate indicator are calculated.

[0036] According to the first-order sensitivity index, all candidate indicators are ranked in descending order. Set a cumulative contribution threshold (such as 85%), select indicators from the top of the list until the sum of the first-order sensitivity indexes of the selected indicators reaches the threshold, and these indicators constitute the "core feature indicator set".

[0037] According to the size of the first-order sensitivity index of each core indicator, an initial weight is assigned to it, which is proportional to the first-order sensitivity index value and is normalized. At the same time, the system identifies typical scenarios such as "production peak period", "device maintenance period", "night quiet period" and other typical scenarios based on historical data clustering analysis, and repeats the analysis and optimization for each scenario to generate a corresponding scenario-specific indicator system.

[0038] Step 5: Update the data samples according to the optimized index system.

[0039] Step 6: Repeat steps 2 to 5 until the preset conditions are met, and obtain the optimized index system for spectrum access decision.

[0040] Specifically, the center node distributes the latest lightweight CNN-Kriging model (obtained by pruning and quantizing the original model) and the scenario-specific indicator system corresponding to the current period (or predicted period) to all edge gateways.​

[0041] The edge gateway side performs the following processes: (1) The edge gateway receives and loads the lightweight model and optimization index system from the center node; (2) When a new Internet of Things device requests access, the gateway obtains its location and service type. For each available candidate channel, the gateway performs the following operations: (a) Call the lightweight CNN-Kriging model, input the device location and candidate channel, and get the baseline interference prediction value ; (b) According to the current local network state (such as the list of connected devices), quickly calculate the real-time values of each core index in the optimization index system. For example, calculate how many devices of the same type are already connected on the current candidate channel; (c) After normalizing the real-time values of each index, perform weighted summation with its dynamic weight to obtain the index-based adjustment term . Weighted summation of the baseline interference prediction value and the index-based adjustment term to obtain the comprehensive interference risk (as an optimization, the weighted coefficients are both set to 0.5); (3) The gateway selects the channel with the smallest comprehensive interference risk value and assigns it to the device requesting access; (4) After the device is connected, the gateway monitors its actual communication quality (such as throughput and retransmission rate) and reports it as new sample data to the center node for iterative updating of the model in the next cycle.

[0042] The above-mentioned index system dynamic construction method based on the CNN-Kriging model is applied to the center node of a distributed intelligent spectrum access system comprising a center node and multiple distributed edge nodes. The method comprises: preprocessing the data samples reported by each distributed edge node, dividing the preprocessing results into a training set and a test set, and training a CNN-Kriging model; generating a synthetic data set according to the trained CNN-Kriging model; statistically analyzing and optimizing the index according to the synthetic data set using the global sensitivity analysis method; updating the data samples according to the optimized index system, updating the index system, and obtaining an optimized index system for spectrum access decision-making. This method solves the problem of low spectrum efficiency, poor reliability, and inability to adapt to dynamic environments caused by static and uninterpretable interference management strategies under the dense deployment of massive Internet of Things devices.

[0043] In one embodiment, step 2 comprises: after normalizing the data samples, obtaining a pre-processing result, dividing the pre-processing result into a training set and a test set; inputting the training samples into the CNN network for feature extraction to obtain a high-dimensional feature vector; taking the high-dimensional feature vector as an input feature of the Kriging module to establish a Kriging surrogate model; the Kriging module is based on Gaussian process regression theory, and a mapping relationship between the input feature and the output response is established by optimizing the covariance function and the hyperparameter; the CNN-Kriging model is trained, cross-validated and regularized based on the training set to obtain a trained CNN-Kriging model; and the trained CNN-Kriging model is tested by using the test set to obtain the trained CNN-Kriging model.

[0044] Specifically, the CNN-Kriging model improves the prediction accuracy and generalization ability of the model by combining the interpolation ability of the Kriging model and the feature extraction and nonlinear fitting ability of the CNN.

[0045] Data preprocessing is the first step of the CNN-Kriging model framework, and its main task is to standardize the original data and divide the data into a training set and a test set to ensure that the subsequent model training and evaluation can be efficiently performed. The fundamental purpose of data preprocessing is to eliminate the dimensional differences between different features and avoid the influence of some features with large value ranges on the model training, thereby ensuring the stability and convergence of the model training. Through normalization, the data is mapped to a unified interval (such as [0, 1] or [-1, 1]), which not only improves the training efficiency of the model, but also effectively prevents problems such as gradient disappearance or gradient explosion, especially in deep learning models, normalization has a significant impact on the convergence speed and final performance of the model. Common normalization methods include Min-Max normalization and Z-score normalization.

[0046] The CNN network is used for feature extraction. The CNN network is one of the core components of the CNN-Kriging model framework. Its main function is to perform multi-level feature extraction on the input data through the CNN network to generate high-dimensional feature vectors, thereby providing high-quality input for subsequent Kriging model construction. The CNN network can effectively capture local features and nonlinear relationships in the data through its unique convolutional layer and pooling layer structure, making it particularly suitable for processing high-dimensional data. The convolutional layer performs local convolution operations on the input data through convolution kernels (filters) to extract spatial or temporal local features of the data. This local connection and weight sharing feature not only significantly reduces the number of model parameters, but also improves computational efficiency. The pooling layer further reduces the spatial dimension of the data through downsampling operations (such as max pooling or average pooling), reducing computational complexity while enhancing the model's robustness to small changes in data. Through the alternating combination of convolutional layers and pooling layers, CNN can gradually extract multi-level features of the data, from low-level edge, texture, and other simple features to high-level features, forming a high-dimensional feature representation with rich expression ability. In this application, the output of the CNN network is a high-dimensional feature vector that not only retains the key information of the original data but also enhances the expression ability of the features through nonlinear transformation, providing more discriminative input for the subsequent Kriging model. This multi-level feature extraction mechanism makes CNN perform well in processing complex data, and CNN can effectively capture the internal structure and patterns of the data, significantly improving the prediction performance of the model. By using the high-dimensional features extracted by CNN as the input of the Kriging model, the hybrid surrogate model can fully utilize the local and nonlinear characteristics of the data while ensuring prediction accuracy, further improving the generalization ability and robustness of the model.

[0047] The Kriging model, as another core component of the CNN-Kriging model framework, is responsible for accurate prediction based on feature selection. The Kriging model is an interpolation method based on Gaussian process theory, and its core idea is to model the spatial correlation between known sample points to predict unknown points. This model not only provides point prediction results but also gives uncertainty estimates of the predicted values. In the hybrid surrogate model framework proposed in this study, the input of the Kriging model is the high-dimensional feature representation generated by the CNN feature extraction module. These features can effectively capture the complex patterns and potential structures of the input data through the multi-layer nonlinear transformation of the convolutional neural network. The Kriging model models these high-dimensional features to establish the mapping relationship between the input features and the objective function, thereby achieving high-precision prediction.

[0048] The construction process of Kriging model involves multiple key steps, among which the selection of correlation function has a great impact on the performance of the model. The correlation function is used to quantify the spatial correlation between sample points, and its selection needs to be combined with the characteristics of the specific problem. Commonly used correlation functions include Gaussian exponential function, cubic spline function, Matern function, etc., which can describe spatial dependence of different scales and strengths. Gaussian exponential function is widely used due to its smoothness and differentiability, while Matern function is more flexible and can adapt to different spatial correlation patterns. After determining the correlation function, the model parameters are usually trained and optimized through optimization methods such as maximum likelihood estimation, and the parameters of the correlation function are determined by maximizing the likelihood function, so that the model can better fit the training data. The prediction results of Kriging model can be combined with other machine learning methods to form a hybrid prediction framework to further improve the performance of the model. In the CNN-Kriging model framework applied for, the construction of Kriging model is closely combined with the CNN feature extraction module to form a complete hybrid surrogate model. By taking the high-dimensional features extracted by CNN as the input of Kriging model, not only can the advantages of CNN in feature extraction be fully utilized, but also the characteristics of Kriging model in high-precision interpolation and uncertainty quantification can be taken advantage of. This combination makes the hybrid surrogate model perform excellently in handling high-dimensional and nonlinear problems.

[0049] The model training and prediction process specifically includes: by integrating the functions of CNN network and Kriging model construction module, a complete hybrid surrogate model system is constructed, which undertakes the key tasks of model optimization and performance evaluation. In the model training phase, the training set data is first processed by the CNN network, which automatically extracts high-dimensional feature representations of the input data through a multi-layer convolutional neural network structure. These features can effectively capture the nonlinear relationship and spatial correlation of the data. Then, the extracted high-dimensional features are input into the Kriging model, which is based on Gaussian process regression theory and establishes the mapping relationship between input features and output responses by optimizing the covariance function and hyperparameters. The training process of Kriging model involves methods such as maximum likelihood estimation to ensure that the model can accurately fit the statistical characteristics of the training data. In the training process, cross-validation, regularization and other means are used to prevent overfitting phenomenon. In the model prediction phase, the test set data is also processed by the CNN feature extraction module to obtain consistent feature representations with the training set. These features are then input into the trained CNN-Kriging model to predict the output response.

[0050] In one embodiment, as Figure 2As shown, the CNN network includes: a first convolutional layer, a first down-sampling layer, a second convolutional layer, a second down-sampling layer, a full connection layer, and an output layer; the output layer includes the full connection layer and a Softmax activation function.

[0051] In one embodiment, the construction process of the Kriging proxy model follows a standard supervised learning process; the construction process of the Kriging proxy model specifically includes: obtaining a set of input-output sample data based on a simulation platform and an engine through a comprehensive performance evaluation method The set of input-output sample data constitutes a training set , wherein is a high-dimensional feature vector, is a corresponding output response calculated through the comprehensive evaluation method; according to the Gaussian process regression theory, a mapping relationship between the input features and the output responses is established by optimizing the covariance function and the hyperparameters, and a Kriging model is obtained; the Kriging model is trained and optimized through a maximum likelihood estimation optimization method.

[0052] Specifically, the construction process of the proxy model follows a standard supervised learning process. Let the original simulation model be , wherein is the dimension of the input parameters. A set of input-output sample data is obtained based on a simulation platform and an engine through a comprehensive performance evaluation method The set of input-output sample data constitutes a training set , wherein is an input vector, is a corresponding output response calculated through the comprehensive evaluation method. The proxy model learns a mapping function through the training set , so that holds for all , wherein is an input space. The training process is completed by minimizing a loss function. Through techniques such as cross-validation, it is ensured that the system performance evaluation proxy model reaches a preset prediction accuracy threshold on an independent test set, so that it can credibly replace the original complex model to perform fast performance deduction and prediction.

[0053] After successfully constructing a high-precision proxy model, a multi-dimensional input space that can comprehensively cover various external operational conditions and internal uncertainty parameters that may be encountered in system operation needs to be designed and defined, , wherein each represents the The value range or probability distribution of each input parameter. By reasonably discretizing or setting the probability distribution of these parameters and considering their possible interaction range, a high-dimensional input parameter space is defined, which can represent the running state of the system in the real complex environment, to ensure that the sensitivity analysis results have practical physical meaning and wide applicability. The dimension of the input space corresponds to all factors that are considered to have an impact on the system effectiveness, which constitutes the initial index pool to be screened. The input parameters need to systematically cover the following aspects: first, the key performance parameters of the system itself, such as the action distance and accuracy of various sensors, the bandwidth and delay of communication links, etc.; second, the organizational structure parameters of the system, such as network topology connection relationship, command level number, information flow path, etc.; third, external environmental conditions; fourth, the statistical characteristics of various uncertainty factors, such as the tolerance range of equipment performance, the accuracy of the fluctuation of operator level, etc. These uncertainties are usually described in the form of probability distribution. By reasonably discretizing or setting the probability distribution of these parameters and considering their possible interaction range, a high-dimensional input parameter space is defined, which can represent the running state of the system in the real complex environment, to ensure that the sensitivity analysis results have practical physical meaning and wide applicability. The dimension of the input space corresponds to all factors that are considered to have an impact on the system effectiveness, which constitutes the initial index pool to be screened. The input parameters need to systematically cover the following aspects: first, the key performance parameters of the system itself, such as the action distance and accuracy of various sensors, the bandwidth and delay of communication links, etc.; second, the organizational structure parameters of the system, such as network topology connection relationship, command level number, information flow path, etc.; third, external environmental conditions; fourth, the statistical characteristics of various uncertainty factors, such as the tolerance range of equipment performance, the accuracy of the fluctuation of operator level, etc. These uncertainties are usually described in the form of probability distribution. By reasonably discretizing or setting the probability distribution of these parameters and considering their possible interaction range, a high-dimensional input parameter space is defined, which can represent the running state of the system in the real complex environment, to ensure that the sensitivity analysis results have practical physical meaning and wide applicability.

[0054] In one embodiment, the correlation function used in the construction process of the Kriging surrogate model is a Gaussian exponential function.

[0055] In one embodiment, the normalization process adopts a minimum-maximum normalization method or a Z-score standardization method.

[0056] Specifically, the minimum-maximum normalization maps the data to a specified interval through linear transformation, and the normalization formula is: ; wherein, represents the normalized value, represents the data to be normalized, and represent the minimum and maximum values of the data, respectively.

[0057] The Z-score standardization converts the data into a distribution with a mean of 0 and a standard deviation of 1 by subtracting the mean and dividing by the standard deviation, and the normalization formula is: ; wherein, represents the normalized value, represents the data to be normalized, and are the mean and standard deviation of the data, respectively.

[0058] Z-score standardization is particularly suitable for cases where data distribution is uneven or there are outliers, which can effectively reduce the interference of outliers on model training. After completing the normalization process, the data is divided into training set and test set, the training set is used for parameter optimization and feature learning of the model, and the test set is used to evaluate the generalization ability and prediction performance of the model. This division method not only can avoid model overfitting, but also can ensure the robustness of the model in actual application. Through scientific data preprocessing, the hybrid proxy model can perform better in subsequent feature extraction and prediction tasks.

[0059] In one embodiment, step 3 comprises: for dimensional input space, generating a Latin hypercube design of sample points, wherein each , and for each dimension , is a random permutation of equiprobable partition of the interval [0, 1]; performing fast calculation on each sampling point by the trained CNN-Kriging model to obtain simulation output data; constructing a synthetic dataset , wherein is the predicted value of the system performance of the th sample point calculated by the trained CNN-Kriging model.

[0060] Specifically, using the constructed high-precision and high-efficiency proxy model, large-scale and randomized sampling is performed in the defined multi-dimensional input space. In order to improve the sampling efficiency, experimental design methods with space-filling characteristics are usually used, such as Latin hypercube sampling. Latin hypercube sampling can achieve more uniform and comprehensive coverage of multi-dimensional parameter space with relatively fewer sample points. Specifically, for dimensional input space, generating a Latin hypercube design of sample points can be expressed as , wherein each , and for each dimension , is a random permutation of equiprobable partition of the interval [0, 1]. Through the proxy model, fast calculation is performed on these sampling points to efficiently generate a large-scale simulation output dataset. This dataset systematically records the corresponding system overall performance output values under different input parameter combinations, forming a sample set , wherein is the predicted value of the system performance of the th sample point calculated by the proxy model.

[0061] In one embodiment, step 4 comprises: decomposing the total variance of the output response into the sum of a first-order effect attributed to the variation of each input parameter alone and second-order and higher-order interaction effects resulting from the interaction between parameters using Sobol variance-based decomposition method; the Sobol decomposition function is: ; wherein, is the output response, is the input random vector, is the constant term, is the first-order effect term, is the second-order interaction effect term, is the d order interaction effect term, d is the number of input parameters.

[0062] The total variance of the output is: ; wherein, is the total variance of the output; is the contribution of the variable to is the joint contribution of the variable and the variable to is the joint contribution of the variable and the variable to is the overall contribution of all to .

[0063] The first-order sensitivity index, the second-order sensitivity index and the total-order sensitivity index are calculated; based on the first-order sensitivity index, the second-order sensitivity index and the total-order sensitivity index and the corresponding preset threshold, the key indicators affecting the system performance are screened out, and the optimized indicator system is obtained.

[0064] Specifically, based on the global sensitivity analysis method, the statistical analysis of the large-scale simulation output data set generated in the previous step is carried out. As a key technology for quantifying the influence of input parameter uncertainty on model output, global sensitivity analysis (GSA) drives the model to perform a large number of simulation calculations by systematically sampling the entire uncertainty space of the input parameters, and then quantifies the overall contribution of each input parameter and its combination to the variance or distribution form of the output result based on the statistical distribution characteristics of the multi-dimensional output response. GSA does not depend on the specific mathematical form of the model, and is especially suitable for handling "black box" or "gray box" models, and can comprehensively analyze the first-order main effect and high-order interaction effect of parameters to obtain more important indicators for the global influence of the system. Usually, Sobol variance decomposition method is used to decompose the total variance of the output response into the first-order effect generated by the change of each input parameter alone, and the sum of the second-order and higher-order interaction effects generated by the interaction between parameters. Based on the variance decomposition principle, the Sobol decomposition decomposes the total variance of the output response into the first-order effect generated by the change of each input parameter alone, and the sum of the second-order and higher-order interaction effects generated by the interaction between parameters. Let the output response , where is the input random vector, and the output is , the total variance of the output is ; where is the constant term, is the first-order effect term, is the second-order interaction effect term, and so on.

[0065] The total variance of the output can be expressed as: ; where , , and so on.

[0066] The Sobol sensitivity index is defined as:

[0067] The first-order sensitivity index measures the contribution proportion of the input parameter to the output variance alone, while the total-order sensitivity index measures the total influence of the parameter and its interaction with all other parameters on the output uncertainty: ; where represents all input parameters except

[0068] By calculating the first-order sensitivity index and the total-order sensitivity index of each parameter, the importance of each indicator can be systematically evaluated. Parameters with high first-order sensitivity index are direct key drivers of system effectiveness; while when the total-order sensitivity index of a parameter is significantly higher than its first-order index, it indicates that the parameter mainly affects the system effectiveness through complex interactions with other parameters. Based on these sensitivity indices, appropriate thresholds can be set to screen out key indicators that have a significant impact on system effectiveness, forming an optimized indicator system, laying the foundation for subsequent effectiveness evaluation and multi-objective optimization.

[0069] It should be understood that although the above Figure 1 steps are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the above Figure 1 steps can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or sub-steps or stages of other steps.

[0070] The embodiment provides a distributed intelligent spectrum access system for implementing the index system dynamic construction method based on the CNN-Kriging model proposed in the application. The system comprises: A central server, the hardware configuration of which includes a high-performance CPU, a GPU, a large-capacity memory and a hard disk. The server runs a model training and optimization software module, a dynamic index system construction software module and a system management interface.

[0071] A plurality of edge Internet of Things gateways are deployed in different areas of the smart farm. Each gateway is equipped with a multi-mode wireless communication unit (such as LoRa, Wi-Fi, BLE), a lightweight computing unit (such as an ARM Cortex-A series processor) and local storage. A local decision engine software runs in the gateway.

[0072] The central server is connected with all edge gateways through a wired or wireless backhaul network. When the system is working, the central server is responsible for performing heavy computing tasks such as model training, sensitivity analysis and index system construction, and issuing the results (lightweight models and index files). Each edge gateway uses the issued resources to independently and in parallel provide low-latency intelligent spectrum access decisions for local devices, and forms a collaborative optimization closed loop with the central server through a feedback mechanism.​

[0073] In one embodiment, a CNN-Kriging model-based index system dynamic construction device is provided, comprising: A data sample acquisition module is configured to acquire multiple groups of data samples reported by each distributed edge node; the data samples include spectrum environment data, device context data, and channel performance data.

[0074] A CNN-Kriging model training module is configured to preprocess the data samples, divide the preprocessing results into a training set and a test set, train the CNN-Kriging model using the training set, then test using the test set, and obtain the trained CNN-Kriging model; the CNN-Kriging model includes a CNN network and a Kriging module; the CNN network is configured to extract features from the training samples, and the Kriging module is configured to construct a Kriging proxy model using the extracted features as input.

[0075] A synthetic data set generation module is configured to generate a synthetic data set through active sampling based on the trained CNN-Kriging model.

[0076] An index system optimization module is configured to statistically analyze and optimize the index system based on the synthetic data set using a global sensitivity analysis method, and determine an optimized index system; and update the data samples based on the optimized index system.

[0077] An optimized index system for spectrum access decision determination module is configured to repeatedly execute the CNN-Kriging model training module, the synthetic data set generation module, and the index system optimization module until a preset condition is met, and obtain an optimized index system for spectrum access decision.

[0078] In one embodiment, the CNN-Kriging model training module is further configured to normalize the data samples to obtain the preprocessing results, divide the preprocessing results into the training set and the test set, input the training samples into the CNN network to extract features and obtain a high-dimensional feature vector, use the high-dimensional feature vector as the input feature of the Kriging module to establish a Kriging proxy model, use the Kriging module based on the Gaussian process regression theory to establish a mapping relationship between the input feature and the output response by optimizing the covariance function and the hyperparameters, train, cross-validate, and regularize the CNN-Kriging model based on the training set to obtain the trained CNN-Kriging model, and test the trained CNN-Kriging model using the test set to obtain the trained CNN-Kriging model.

[0079] In one embodiment, the CNN network in the CNN-Kriging model training module comprises a first convolutional layer, a first down-sampling layer, a second convolutional layer, a second down-sampling layer, a full connection layer, and an output layer; the output layer comprises the full connection layer and a Softmax activation function.

[0080] In one embodiment, the construction process of the Kriging surrogate model in the CNN-Kriging model training module follows a standard supervised learning process; the construction process of the Kriging surrogate model specifically comprises: obtaining a set of input-output sample data based on a simulation platform and an engine through a comprehensive performance evaluation method The set of input-output sample data constitutes a training set , wherein is a high-dimensional feature vector, is a corresponding output response calculated through the comprehensive evaluation method; according to the Gaussian process regression theory, a mapping relationship between the input features and the output responses is established by optimizing the covariance function and the hyperparameters, to obtain a Kriging model; the Kriging model is trained and optimized through a maximum likelihood estimation optimization method.

[0081] In one embodiment, the correlation function used in the construction process of the Kriging surrogate model in the CNN-Kriging model training module is a Gaussian exponential function.

[0082] In one embodiment, the normalization processing in the CNN-Kriging model training module adopts a min-max normalization method or a Z-score standardization method.

[0083] In one embodiment, the synthetic data set generation module is further configured to generate a Latin hypercube design of sample points for a -dimensional input space. , wherein each , and for each dimension , is a random arrangement of equal-probability partitioning of the interval [0, 1]; the trained CNN-Kriging model is used to quickly calculate each sampling point to obtain simulation output data; all sampling points and corresponding simulation output data constitute a synthetic data set , wherein is a system performance prediction value of the th sample point calculated through the trained CNN-Kriging model.

[0084] In one embodiment, the index system optimization module is further configured to decompose a total variance of the output response into a sum of a first-order effect attributed to a change of each input parameter alone and a second-order and higher-order interaction effect attributed to an interaction between parameters by using a Sobol variance decomposition method; the Sobol decomposition function is: ; wherein, is the output response, is an input random vector, is a constant term, is a first-order effect term, is a second-order interaction effect term, is an d order interaction effect term, d is a number of input parameters; a total variance of the output is: ; wherein, is the total variance of the output; is a contribution of a variable to is a joint contribution of a variable and a variable to is a joint contribution of a variable and a variable to is an overall contribution of all variables to .

[0085] The first-order sensitivity index, the second-order sensitivity index, and the total-order sensitivity index are calculated; based on the first-order sensitivity index, the second-order sensitivity index, and the total-order sensitivity index and corresponding preset thresholds, key indicators affecting the system performance are screened, and an optimized index system is obtained.

[0086] It can be understood that the specific explanations and descriptions of the index system dynamic construction device based on the CNN-Kriging model can refer to the corresponding explanations and descriptions of the embodiments of the index system dynamic construction method based on the CNN-Kriging model described above, and will not be repeated here. Each module in the above index system dynamic construction device based on the CNN-Kriging model can be realized by software, hardware, and combinations thereof, in whole or in part. The above modules can be embedded in or independent of a device with data processing function in hardware form, or can be stored in the memory of the aforementioned device in software form, so as to call and execute the operations corresponding to each module by the processor. The aforementioned device can be, but is not limited to, various types of data processing computer devices in the prior art.

[0087] Any technical features in the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described above, however, any combination of the technical features is considered to be within the scope of the present disclosure.

[0088] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the protection scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, some modifications and improvements can be made without departing from the concept of the present application, and all of them belong to the protection scope of the present application.

Claims

1. A method for dynamically constructing an indicator system based on a CNN-Kriging model, characterized in that, This method is applied to the central node of a distributed intelligent spectrum access system that includes a central node and multiple distributed edge nodes, and includes the following steps: Step 1: Obtain multiple sets of data samples reported by each distributed edge node; the data samples include spectrum environment data, device context data, and channel performance data; Step 2: Preprocess the data samples, dividing the preprocessing results into a training set and a test set; train the CNN-Kriging model using the training set, and then test it using the test set to obtain the trained CNN-Kriging model; the CNN-Kriging model includes: a CNN network and a Kriging module; the CNN network is used to extract features from the training samples, and the Kriging module is used to construct a Kriging surrogate model using the extracted features as input; Step 3: Generate a synthetic dataset based on the trained CNN-Kriging model through active sampling; Step 4: Based on the synthetic dataset, perform statistical analysis and index optimization using the global sensitivity analysis method to determine the optimized index system; Step 5: Update the data sample according to the optimized indicator system; Step 6: Repeat steps 2 to 5 until the preset conditions are met to obtain the optimized index system for spectrum access decision-making.

2. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 1, characterized in that, Step 2 includes: After normalizing the data samples, a preprocessing result is obtained, which is then divided into a training set and a test set. The training samples are input into the CNN network for feature extraction to obtain high-dimensional feature vectors; The high-dimensional feature vector is used as the input feature of the Kriging module to establish a Kriging surrogate model; the Kriging module is based on Gaussian process regression theory and establishes the mapping relationship between input features and output response by optimizing the covariance function and hyperparameters; The CNN-Kriging model is trained, cross-validated, and regularized based on the training set to obtain the trained CNN-Kriging model; The trained CNN-Kriging model was tested using a test set to obtain the well-trained CNN-Kriging model.

3. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 2, characterized in that, The CNN network includes: a first convolutional layer, a first downsampling layer, a second convolutional layer, a second downsampling layer, a fully connected layer, and an output layer; the output layer includes a fully connected layer and a Softmax activation function.

4. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 2, characterized in that, The construction process of the Kriging agent model follows the standard supervised learning process; the specific steps in the construction process of the Kriging agent model include: Based on the simulation platform and engine, a comprehensive performance evaluation method is used to obtain... The training set consists of input-output sample data. ,in It is a high-dimensional feature vector. The corresponding output response is calculated using a comprehensive evaluation method; Based on the Gaussian process regression theory, the Kriging model is obtained by optimizing the covariance function and hyperparameters to establish the mapping relationship between input features and output response. The Kriging model is trained and optimized using the maximum likelihood estimation optimization method.

5. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 4, characterized in that, The relevant function used in the construction of the Kriging proxy model is the Gaussian exponential function.

6. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 2, characterized in that, Normalization is performed using either the min-max normalization method or the Z-score standardization method.

7. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 1, characterized in that, Step 3 includes: for 3D input space, generate Latin hypercube design for each sample point Each of them And for each dimension , It is a random permutation of the interval [0,1] with equal probability; The simulation output data is obtained by quickly calculating each sampling point using a pre-trained CNN-Kriging model. All sampling points and the corresponding simulation output data system constitute a synthetic dataset. ,in The first one is calculated using a trained CNN-Kriging model. Predicted system performance values ​​for each sample point.

8. The method for dynamically constructing an indicator system based on the CNN-Kriging model according to claim 1, characterized in that, Step 4 includes: The Sobol variance decomposition method is used to decompose the total variance of the output response into the sum of first-order effects attributable to the individual variations of each input parameter, and second-order and higher-order interaction effects resulting from the interactions between parameters; the Sobol decomposition function is: in, To output the response, Given a random input vector, For constant terms, This is a first-order effect term. This is a second-order interaction effect term. for d The first-order interaction term, d The number of input parameters; The total output variance is: in, For output The total variance; For variables right Contribution For variables and variables right The joint contribution For all The overall contribution; Calculate the first-order sensitivity index, the second-order sensitivity index, and the overall sensitivity index; Based on the first-order sensitivity index, the second-order sensitivity index, and the total-order sensitivity index, as well as the corresponding preset thresholds, key indicators affecting the system performance are selected, resulting in an optimized indicator system.

9. A dynamic indicator system construction device based on the CNN-Kriging model, characterized in that, include: The data sample acquisition module is used to acquire multiple sets of data samples reported by each distributed edge node; the data samples include spectrum environment data, device context data, and channel performance data. The CNN-Kriging model training module is used to preprocess the data samples and divide the preprocessing results into training set and test set; The CNN-Kriging model is trained using a training set and then tested using a test set to obtain the trained CNN-Kriging model. The CNN-Kriging model includes a CNN network and a Kriging module. The CNN network is used to extract features from the training samples, and the Kriging module is used to construct a Kriging surrogate model using the extracted features as input. The synthetic dataset generation module is used to generate synthetic datasets based on the trained CNN-Kriging model through active sampling. The indicator system optimization module is used to perform statistical analysis and indicator optimization based on the synthetic dataset using global sensitivity analysis, to determine the optimized indicator system, and to update the data samples based on the optimized indicator system. The module for determining the optimal index system for spectrum access decision-making is used to repeatedly execute the CNN-Kriging model training module, the synthetic dataset generation module, and the index system optimization module until the preset conditions are met, thereby obtaining the optimal index system for spectrum access decision-making.

Images