Endogenous intelligent system based on wireless data knowledge graph and construction method thereof

Abstract

The application provides an endogenous intelligent system and a construction method based on a wireless data knowledge graph, and the system comprises a wireless data knowledge graph intelligent generation unit, a feature data set generation unit, a wireless network digital twin unit and an AI model unit. The wireless data knowledge graph intelligent generation unit generates a wireless data knowledge graph by using a preset neural network algorithm on partial wireless communication data in a business scenario; a key feature data set of a target business is extracted from the wireless data knowledge graph according to a feature data set generation algorithm; an initial target business strategy is generated by the AI model unit according to the key feature data set, and the strategy is pre-verified by the wireless network digital twin unit; and the pre-verification result is fed back to the AI model unit, and the final target business strategy is issued after adjustment according to the pre-verification feedback result. The application reduces the cost related to data collection and calculation, and supports the realization of real-time and green network intelligence.

Description

Technical Field

[0001] The present invention relates to the field of intelligent wireless communication network technology, and in particular to an endogenous intelligent system based on a wireless data knowledge graph and a construction method. Background Art

[0002] The endogenous AI (artificial intelligence) built into green communication systems must meet stringent real-time requirements. To achieve this, deploying lightweight and resource-efficient AI models is essential. However, wireless networks generate a vast array of data fields and metrics during service interactions, only a small portion of which significantly impacts the endogenous AI models. Therefore, the real-time intelligence of communication systems relies primarily on a small but crucial set of data, which profoundly influences the performance of the AI ​​models.

[0003] Therefore, how to achieve lightweight network-built-in AI models using a small amount of data is an urgent problem that needs to be solved. Summary of the Invention

[0004] The present invention provides an endogenous intelligent system and construction method based on wireless data knowledge graph, which is used to solve the defects of the existing technology in that communication data in the network is redundant, resulting in the inability of AI models to be lightweight and to process data in real time. It realizes the lightweighting of the network's built-in AI model through a small amount of data.

[0005] The present invention provides an endogenous intelligent system based on a wireless data knowledge graph, comprising: a wireless data knowledge graph intelligent generation unit, a feature data set generation unit, a wireless network digital twin unit, and an AI model unit;

[0006] The wireless data knowledge graph intelligent generation unit is used to input local wireless communication data in the business scenario into a preset neural network model to generate a target wireless data knowledge graph model;

[0007] The feature data set generation unit is used to extract a key feature data set of the target service in real time from the target wireless data knowledge graph model according to a feature data set generation algorithm; the key feature data set is a valid data set associated with the target service;

[0008] The AI ​​model unit is used to generate an initial target service policy based on the key feature data set, and send the initial target service policy to the wireless network digital twin unit for policy pre-verification;

[0009] The wireless network digital twin unit is configured to perform pre-verification of a policy according to the key feature data set and the initial target service policy, and feed back the pre-verification result to the AI model unit, so that the AI model unit adjusts and issues a final target service policy according to the pre-verification feedback result.

[0010] According to the application, a wireless data knowledge graph-based endogenous intelligent system is provided, and the wireless data knowledge graph intelligent generation unit includes a non-real-time data acquisition module, a data preprocessing module and a knowledge graph generation module.

[0011] The non-real-time data acquisition module is configured to acquire multiple types of non-real-time data in a service scenario by one or a combination of multiple modes of hard acquisition, soft acquisition or road testing.

[0012] The data preprocessing module is configured to perform one or more operations of deleting, time distinguishing and structuring on the multiple types of non-real-time data to obtain preprocessed data.

[0013] The knowledge graph generation module is configured to construct a local wireless data knowledge graph model according to the preprocessed data, and generate a target wireless data knowledge graph model according to the local wireless data knowledge graph model and a preset neural network model.

[0014] According to the application, a wireless data knowledge graph-based endogenous intelligent system is provided, and the feature data set generation unit includes a feature sorting module, a feature data set generation module and a feature data set evaluation module,

[0015] The feature sorting module is configured to sort features according to the influence degree of other nodes on a target KPI node in the target wireless data knowledge graph model.

[0016] The feature data set generation module is configured to determine a key feature data set according to the sorting result and a feature data set generation algorithm.

[0017] The feature data set evaluation module is configured to evaluate the key feature data set.

[0018] The application further provides a wireless data knowledge graph-based endogenous intelligent system construction method, which includes:

[0019] Generating a target wireless data knowledge graph model based on local wireless communication data in a service scenario and a preset neural network model;

[0020] Extracting a key feature data set of a target service in the target wireless data knowledge graph model according to a feature data set generation algorithm; the key feature data set is a minimum amount of effective data set associated with the target service;

[0021] Inputting the key feature dataset into the AI ​​model to generate an initial target business strategy;

[0022] Inputting the initial target service strategy and the key feature data set into the wireless network digital twin for pre-verification;

[0023] Feeding pre-verification results back to the AI ​​model;

[0024] Adjust the initial target business strategy based on the pre-verification feedback results until the final target business strategy is issued.

[0025] According to a method for constructing an endogenous intelligent system based on a wireless data knowledge graph provided by the present invention, generating a target wireless data knowledge graph model based on local wireless communication data in a business scenario and a preset neural network model includes:

[0026] Constructing a local wireless data knowledge graph model; the local wireless data knowledge graph model at least includes a real adjacency matrix and a real data matrix of each node;

[0027] Performing standardization processing on the real data matrix of each node to obtain a standardized real data matrix of each node;

[0028] Generate a training set based on the standardized real data matrix of each node;

[0029] The training set is input into a preset neural network model for training to obtain a target wireless data knowledge graph model.

[0030] According to a method for constructing an endogenous intelligent system based on a wireless data knowledge graph provided by the present invention, inputting the training set into a preset neural network model for training to obtain a target wireless data knowledge graph model includes:

[0031] Inputting the training set into a preset neural network model for training to obtain a node embedding matrix;

[0032] Generate a reconstructed adjacency matrix according to the node embedding matrix and a preset reconstruction matrix generation algorithm;

[0033] Calculating the loss value of the spatiotemporal heterogeneous graph attention neural network model using the reconstructed adjacency matrix and the true adjacency matrix;

[0034] Determine whether the loss value meets the preset loss condition. If so, obtain the target wireless data knowledge graph model; otherwise, continue training.

[0035] According to a method for constructing an endogenous intelligent system based on a wireless data knowledge graph provided by the present invention, the method further includes:

[0036] Dividing the local wireless data knowledge graph model into multiple graph slice models according to coherence time;

[0037] Correspondingly, performing normalization processing on the real data matrix of each node to obtain the normalized real data matrix of each node includes:

[0038] The node real data matrix corresponding to each graph slice model is standardized to obtain the standardized real data matrix of each node.

[0039] According to a method for constructing an endogenous intelligent system based on a wireless data knowledge graph provided by the present invention, extracting a key feature dataset of a target business from the wireless data knowledge graph according to a feature dataset generation algorithm includes:

[0040] Obtain the influence set of each node in the target wireless data knowledge graph model on the target KPI (Key Performance Indicators) node;

[0041] Sorting the nodes in the influence set;

[0042] The node feature with the highest influence ranking is used as the dependent variable and input into the KPI prediction algorithm to obtain the KPI prediction value;

[0043] Calculate the degree of fit based on the KPI predicted value, the actual KPI value and the degree of fit calculation rule;

[0044] comparing the fitness with a preset fitness;

[0045] If the degree of fit is less than the preset degree of fit, then continue to add the feature of the next node in the influence set as a common dependent variable, and recalculate the KPI prediction value and the degree of fit until it is equal to or greater than the preset degree of fit;

[0046] Generate key feature data sets for the target business based on the nodes that have achieved the preset fit and the real data collected by the nodes.

[0047] According to a method for constructing an endogenous intelligent system based on a wireless data knowledge graph provided by the present invention, the method further includes:

[0048] The feature compression rate is calculated based on the number of extracted nodes that reach the fitting degree and the total number of nodes in the wireless data knowledge graph;

[0049] The key feature dataset is evaluated using the feature compression ratio.

[0050] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above methods when executing the program.

[0051] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which implements any of the above methods when executed by a processor.

[0052] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any one of the above methods.

[0053] The present invention provides an endogenous intelligent system and construction method based on a wireless data knowledge graph. By utilizing a wireless data knowledge graph and a method for generating a feature dataset, the endogenous intelligent system is constructed. Large amounts of complex data are collected from wireless communication networks to construct a wireless data knowledge graph, and key feature datasets that influence target services are screened from the wireless data knowledge graph. The AI ​​model drives real-time AI training and inference by collecting a significantly reduced feature dataset in real time, thereby achieving efficient, real-time endogenous intelligence in wireless networks. The present invention only requires the real-time collection of a small number of key data fields to train lightweight AI models, thereby reducing the costs associated with data collection and computation and enabling the realization of real-time, green network intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 This is a schematic diagram of the structure of the endogenous intelligent system based on the wireless data knowledge graph provided by the present invention;

[0056] Figure 2 This is a schematic diagram of the interaction within the endogenous intelligent system based on the wireless data knowledge graph provided by the present invention;

[0057] Figure 3 This is a flow chart of a method for constructing an endogenous intelligent system based on a wireless data knowledge graph provided by the present invention;

[0058] Figure 4 It is a flow chart of the method for constructing a wireless data knowledge graph provided by the present invention;

[0059] Figure 5This is a schematic diagram of the spatiotemporal heterogeneous graph attention neural network model provided by the present invention;

[0060] Figure 6 This is a schematic diagram of generalized element path division provided by the present invention;

[0061] Figure 7 It is a schematic diagram of a graph slice within the coherence time provided by the present invention;

[0062] Figure 8 It is a flowchart of the method for generating a feature data set provided by the present invention;

[0063] Figure 9 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0065] The following combination Figure 1-Figure 2 The endogenous intelligent system based on wireless data knowledge graph of the present invention is described, comprising: a wireless data knowledge graph intelligent generation unit 101, a feature data set generation unit 102, a wireless network digital twin unit 103 and an AI model unit 104;

[0066] The wireless data knowledge graph intelligent generation unit 101 is used to input local wireless communication data in a business scenario into a preset neural network model to generate a target wireless data knowledge graph model.

[0067] In this embodiment, full-service wireless communication data is collected in a non-real-time manner. A wireless data knowledge graph is learned and constructed semi-dynamically to analyze, understand, and represent the intrinsic relationships between current data fields.

[0068] Local wireless communication data refers to some data fields related to the service.

[0069] The feature data set generation unit 102 is used to extract a key feature data set in real time from the target wireless data knowledge graph model according to a feature data set generation algorithm; the key feature data set is the minimum valid data set associated with the target service; the minimum valid data set is obtained according to the feature data set generation algorithm.

[0070] In this embodiment, the key feature data set is a feature data set with a large number of features, that is, a minimum valid data set. The minimum valid data set is important data that has a great impact on the target business.

[0071] The AI ​​model unit 104 is used to generate an initial target business policy based on the key feature data set, and send the initial target business policy to the wireless network digital twin unit for policy pre-verification.

[0072] In this embodiment, the AI ​​model unit 104 performs training and reasoning by collecting key feature data sets and issuing strategies in real time, thereby realizing efficient real-time endogenous intelligence of the wireless network.

[0073] The wireless network digital twin unit 105 is used to perform policy pre-verification based on the key feature data set and the initial target business policy, and feed back the pre-verification result to the AI ​​model unit 104, so that the AI ​​model unit 104 can adjust the final target business policy based on the pre-verification feedback result.

[0074] In this embodiment, key feature data sets are used to provide feature references and data inputs for wireless network digital twin units to accurately reflect the virtual model of the real physical network and provide policy pre-verification functions in the digital world to support decision makers in making more reasonable and correct network resource management strategies and achieve optimization of the real physical network.

[0075] The wireless network digital twin unit 105 is a pre-verification function for the resource management and control strategy of the AI ​​model unit 104. It verifies whether the strategy issued by the AI ​​model is feasible and provides the AI ​​model with strategy generation opinions. The AI ​​model can adjust the strategy based on the pre-verification feedback results until the preset effect is achieved and then issue the final target business strategy.

[0076] In a specific embodiment, the target service is uplink throughput. A total of 188 data fields of real data related to uplink throughput are collected, from which 82 data fields related to uplink throughput are selected to construct a wireless data knowledge graph. Each data field is a node in the graph, and a total of 120,000 data are collected for each node. The feature data set generation algorithm is used to extract the key feature data set of uplink throughput from the wireless data knowledge graph. A total of 4 data fields, corresponding to four nodes, are extracted. The data of these 4 data fields are input into the AI ​​model to obtain the uplink throughput. The uplink throughput is input into the wireless network digital twin to verify whether the throughput is abnormal, and the verification result is fed back to the AI ​​model.

[0077] Experimental results show that when all 188 node data is fed into the AI ​​model, the throughput fit achieved is 99%, with a computation time of 465.75 seconds and a floating-point computation load of 1.63*10^(-5)G. When data from four nodes is fed into the AI ​​model, the throughput fit achieved is 97%, with a computation time of 28.33 seconds and a floating-point computation load of 4.41*10^(-6)G. Therefore, the uplink throughput predicted by feeding the data from four nodes into the AI ​​model can be used as the basis for the final strategy. This runtime is reduced by more than 1 / 10 compared to the method used on the original dataset, reducing training overhead by an order of magnitude.

[0078] This system architecture only requires real-time collection of a small number of key data fields to train lightweight AI models, significantly reducing resource waste and overhead in the communication network and supporting the realization of real-time, green network intelligence.

[0079] In one embodiment, the wireless data knowledge graph intelligent generation unit 101 includes a non-real-time data acquisition module, a data preprocessing module and a knowledge graph generation module;

[0080] The non-real-time data acquisition module is used to collect multiple types of non-real-time data in business scenarios through a combination of one or more of hard acquisition, soft acquisition or drive testing.

[0081] In an example, the multiple types of non-real-time data mainly include one or more of UE-side wireless air interface data, base station-side wireless air interface data, core network data, and network management data.

[0082] The data preprocessing module performs one or more operations of deleting, distinguishing by time, and structurally processing the multiple types of non-real-time data to obtain preprocessed data.

[0083] The knowledge graph generation module is used to construct a local wireless data knowledge graph model based on the preprocessed data, and generate a target wireless data knowledge graph model based on the local wireless data knowledge graph model and a preset neural network model.

[0084] In a specific embodiment, for example, when monitoring downlink throughput, 200 data fields related to downlink throughput are first collected. 100 data fields are randomly selected from these 200 data fields and used to construct a local downlink throughput knowledge graph. Subsequently, dirty data in the 200 data fields related to downlink throughput is deleted. After deletion, the data is partitioned by time and structured. Using the data from the processed 200 data fields related to downlink throughput and the constructed local downlink throughput knowledge graph, a spatiotemporal heterogeneous graph attention neural network model is used for continuous analysis and mining to complete the downlink throughput knowledge graph.

[0085] The present invention generates business-related knowledge graphs through a wireless data knowledge graph intelligent generation unit, which not only enriches business data, but also clearly and intuitively obtains the relationship between various data fields, providing a rich and effective data support for the AI ​​model.

[0086] In the embodiment of the present invention, the feature data set generating unit 102 further includes a feature sorting module, a feature data set generating module and a feature data set evaluating module.

[0087] The feature sorting module is used to perform feature sorting according to the influence of other nodes in the target wireless data knowledge graph model on the target KPI node.

[0088] By sorting features, data that is important to the target KPI node can be placed at the front, making it easier to extract key feature data.

[0089] The feature data set generation module is used to determine the key feature data set according to the sorting result and the feature data set generation algorithm.

[0090] The feature data set evaluation module is used to evaluate the key feature data set.

[0091] Through evaluation, we can further determine whether the key feature dataset can achieve the same effect as the original data.

[0092] In a specific embodiment, for example, for monitoring uplink throughput, a wireless data knowledge graph is constructed based on 82 data fields, each data field corresponds to a node, one of the nodes is selected as the target KPI node, the influence of other nodes on the target KPI node is calculated, and the other nodes are sorted according to the influence. The first four nodes and the real data obtained by the nodes are determined through the feature data set generation algorithm to generate a key feature data set, and the feature data set is evaluated.

[0093] The present invention uses the feature data set generation unit 102 to extract a small amount of key data from massive and complex data, and trains and infers the AI ​​model through a small amount of key data, which greatly reduces computing costs, saves computing power, and ensures the efficiency of real-time communication in the 6G network.

[0094] Among them, in the system of this embodiment, the non-real-time data acquisition module, the data preprocessing module, the knowledge graph generation module, the feature sorting module, the feature data set generation module, the feature data set evaluation module, and the AI ​​model unit 104 sequentially form a non-real-time outer loop communication network structure, and the data loop of non-real-time wireless data acquisition, key feature data set extraction and policy issuance is realized through the non-real-time outer loop communication network structure.

[0095] The feature sorting module, feature data set generation module, feature data set evaluation module, wireless network digital twin unit 105 and AI model unit 104 sequentially form a real-time inner loop communication network structure. Through the real-time inner loop communication structure, the inner data loop of feature data collection, policy issuance and policy verification is realized.

[0096] At the outer layer, wireless communication data is collected in a non-real-time manner. A wireless data knowledge graph is semi-dynamically learned and constructed to analyze, understand, and represent the inherent relationships between current data fields. This allows for the selection of key feature datasets. Guided by the outer layer, the inner layer collects a significantly reduced key feature dataset in real time and drives AI model training and inference, thus achieving efficient, real-time, endogenous intelligence within the wireless network.

[0097] In a specific embodiment, Figure 2 As shown in the figure, the interaction process of each module is as follows:

[0098] In the first step S201, a non-real-time data collection module collects multiple types of non-real-time data;

[0099] In the second step S202, the data preprocessing module performs operations such as deletion and structural processing on multiple types of non-real-time data; the knowledge graph generation module generates a wireless data knowledge graph based on the preprocessed data in the second step;

[0100] In the third step S203, the feature sorting module sorts the standardized real data in the wireless data knowledge graph;

[0101] In the fourth step S204, the sorted features are input into the feature dataset generation and evaluation module to generate a key feature dataset;

[0102] In the fifth step S205, the key feature data set is sent to the Na Yuan and real-time data acquisition modules of the wireless network digital twin respectively;

[0103] In step S206, the real-time data collection module inputs the collected key feature data into the real-time strategy delivery module;

[0104] In the seventh step S207, the real-time policy issuing module sends the generated policy to the wireless network digital twin unit for pre-verification;

[0105] In the eighth step S208, the wireless network digital twin unit feeds back the pre-verification result to the real-time strategy issuing module.

[0106] like Figure 3 As shown, the present invention also provides an endogenous intelligence method based on wireless data knowledge graph, comprising the following steps:

[0107] Step S301: Generate a target wireless data knowledge graph model based on local wireless communication data in the business scenario and a preset neural network model;

[0108] Step S302: extracting a key feature dataset in real time from the target wireless data knowledge graph model according to a feature dataset generation algorithm; the key feature dataset is a valid dataset associated with the target service;

[0109] Step S303: Inputting the key feature data set into the AI ​​model to generate an initial target business strategy;

[0110] Step S304: inputting the initial target service strategy and the key feature data set into the wireless network digital twin for pre-verification;

[0111] Step S305: Feedback the pre-verification result to the AI ​​model;

[0112] Step S306: Adjust the initial target service policy according to the pre-verification feedback result until the final target service policy is issued.

[0113] This embodiment refers to a specific embodiment of the endogenous intelligent system. For example, the target service is uplink throughput. A total of 188 data fields of real data related to uplink throughput are collected, and 82 data fields related to uplink throughput are selected to construct a wireless data knowledge graph. Each data field is a node in the graph, and a total of 120,000 data are collected for each node. The feature data set generation algorithm is used to extract the key feature data set of uplink throughput from the wireless data knowledge graph. A total of 4 data fields, corresponding to four nodes, are extracted. The data of these 4 data fields are input into the AI ​​model to obtain the uplink throughput. The uplink throughput is then input into the wireless network digital twin to verify whether the throughput is abnormal, and the verification result is fed back to the AI ​​model.

[0114] Experimental results show that when all 188 node data is fed into the AI ​​model, the throughput fit achieved is 99%, with a computation time of 465.75 seconds and a floating-point computation load of 1.63*10^(-5)G. When data from four nodes is fed into the AI ​​model, the throughput fit achieved is 97%, with a computation time of 28.33 seconds and a floating-point computation load of 4.41*10^(-6)G. Therefore, the uplink throughput predicted by feeding the data from four nodes into the AI ​​model can be used as the basis for the final strategy. This runtime is reduced by more than 1 / 10 compared to the method used on the original dataset, reducing training overhead by an order of magnitude.

[0115] Through the embodiment of the present invention, based on the wireless data knowledge graph model, only a small amount of key data fields need to be collected in real time to train a lightweight AI model, which significantly reduces the resource waste and overhead of the communication network and supports the realization of real-time, green network intelligence.

[0116] like Figure 4 As shown, the target wireless data knowledge graph model is generated based on the local wireless communication data in the business scenario and the preset neural network model, including:

[0117] Step S401: Construct a local wireless data knowledge graph model; the local wireless data knowledge graph model includes a real adjacency matrix and a real data matrix of each node.

[0118] In this embodiment, a local wireless data knowledge graph model can be constructed by collecting historical business data and extracting some relevant data from the historical business data.

[0119] In a specific example, for example, the constructed wireless data knowledge graph is recorded as in, Represents the wireless data knowledge graph. express The total number of nodes in the set is N. ε represents The set of all edges in , where edge e ij Type of (R) i,j express, Represents the set of all edge relationship types. W is a node static attribute matrix, which represents the fixed and statistical attributes associated with each node. Each row of W represents a node, and the columns represent F attributes. The attributes are mainly determined by the protocol, such as node type, communication layer, and adjustability. T is the sampling time matrix; A represents the adjacency matrix of the wireless data knowledge graph, which is represented by (A) i,j Represents the element in row i and column j of A. If there is an edge between node i and node j, then (A) i,j If , it is 1, otherwise it is 0. X represents the matrix formed by the real data collected by each node in the wireless data knowledge graph.

[0120] Step S402: performing standardization processing on the real data matrix of each node to obtain a standardized real data matrix of each node.

[0121] In this embodiment, since the original data matrix is ​​missing values ​​at certain moments during the actual acquisition process, it is first necessary to use linear interpolation to fill in the missing values ​​of X, and then calculate the mean and variance values, and perform normalization based on the mean and variance values.

[0122] In a specific example, suppose the data If there is a gap, it is replaced by the mean of the previous and next moments:

[0123]

[0124] in, is the i-th data collected at time t, is the data collected at the previous moment, is the i-th data collected at the next moment.

[0125] Since the value ranges of each field fluctuate greatly, it is also necessary to perform z-score standardization on the completed data matrix. Find the mean μ u and variance σ i , then for any The standardized value is calculated according to the following formula

[0126]

[0127] Depend on The standardized data matrix composed of

[0128] Step S403: Generate a training set by normalizing the real data matrix of each node.

[0129] In this embodiment, when dividing the training set and the test set, the total number of nodes in the training set is first calculated to be γ·N, and the total number of nodes in the test set is (1-γ)·N. Randomly select γ·N nodes from the dataset, and the real data matrix of the selected nodes is used as the training set for training. The real data matrix of the remaining (1-γ)·N nodes is used as the test set. The training set nodes after each random division form a new graph.

[0130] Step S404: input the training set into a preset neural network model for training to obtain a target wireless data knowledge graph model.

[0131] In an embodiment of the present invention, inputting the training set into a preset neural network model for training to obtain a target wireless data knowledge graph model includes:

[0132] Step S404a: inputting the training set into a preset neural network model for training to obtain a node embedding matrix.

[0133] In this application, the preset neural network can adopt a spatiotemporal heterogeneous graph attention neural network.

[0134] like Figure 5As shown in the figure, the spatiotemporal heterogeneous graph attention neural network of this method includes two convolutional layers, namely the temporal convolutional layer and the heterogeneous graph attention layer, in addition to the input layer and the output layer.

[0135] In one example, the spatiotemporal heterogeneous graph attention neural network can adopt the STREAM embedding learning model, which can accurately capture the degree of correlation between wireless data fields and graph structures by efficiently extracting the spatial, temporal and attribute information of wireless data, thereby intelligently generating a wireless data knowledge graph.

[0136] In one embodiment, before training, the network parameters of the spatiotemporal heterogeneous graph attention neural network are first initialized. The parameters include: the ratio of nodes in the training set to all nodes γ, the number of times new graphs are generated G, the number of training rounds for each graph E, and the loss function threshold and the learning rate r.

[0137] In the temporal convolutional layer, Represents the cth convolution kernel of the temporal convolution layer, where K S , represents the dimension of the convolution kernel in terms of the number of features; K T represents the dimension of the convolution kernel in time, c in Indicates the dimension of the convolution kernel on the input channel; represents the total dimension of the convolution kernel, is a real number. Then the convolution kernel is the feature tensor of the lth layer The calculation formula for time convolution is:

[0138]

[0139] in,(·) n,m,c means taking the (n,m,c)th element of the matrix or tensor, so express The (n,m,c)th element of , σ represents the activation function, Represents the temporal convolution operation. c in and c out Represent the dimensions of input channels and output channels respectively. The lth layer has c out Convolution kernels are used to construct the feature tensor of the l+1th layer by splicing the results of these convolution kernels together:

[0140]

[0141] The feature tensor of each layer in the graph convolution layer can be calculated by performing convolution operations layer by layer through Formula 4. Among them, the feature tensor of the first layer is the real matrix X in the training set.

[0142] In the heterogeneous graph attention layer, there are node-level attention aggregation and meta-path-level attention aggregation. The specific steps are as follows:

[0143] The first step is to calculate the graph convolution, which specifically includes:

[0144] set up is a graph convolution kernel. Combined with the activation function σ, the multi-layer propagation rule of GCN (Graph Convolutional Networks) is expressed as follows:

[0145]

[0146] Where O is the training parameter matrix, Represents the graph convolution operation; l represents any layer in the convolution layer; To normalize the adjacency matrix, it is calculated using the following formula:

[0147]

[0148] Where I represents the identity matrix.

[0149] is the normalized degree matrix, D is the degree matrix of the wireless knowledge graph, which is a diagonal matrix, and its i-th diagonal element is defined as Similarly, the normalized degree matrix is Its i-th diagonal element is defined as

[0150] The convolution result of the jth kernel can be calculated using the following formula:

[0151]

[0152] where c in and c out Represent the dimensions of input channels and output channels respectively. That is, there are c out kernels participate in the graph convolution of the lth layer.

[0153] Similarly, the convolution operation can be performed layer by layer through Formula 9 to calculate the representation tensor of each layer in the heterogeneous graph attention layer. Among them, the representation tensor of the first layer is equal to the real data matrix X, that is,

[0154] The second step is to perform node-level attention aggregation operations including:

[0155] First, the generalized meta-path set obtained based on the edge relationship categories of the wireless data knowledge graph specifically includes:

[0156] Generalized metapath φ p yes The abbreviations R1, R2, R3, (·) indicate the type of any node. Figure 6 As shown in Figure 2, the definition of generalized meta-path can be used to divide the wireless data knowledge graph into three sub-graphs according to edge categories. Here, the generalized meta-path set can be directly obtained according to the definition of generalized meta-path.

[0157] For a given metapath φ p Node pairs on (υ i ,υ j ), node-level attention coefficient from node i to node j The calculation is as follows:

[0158]

[0159] Among them, || represents the vector splicing operation, represents the embedding matrix of node i in the l-th spatial convolutional layer, Represents the meta-path φ p The node-level attention vector of .

[0160] After obtaining the attention coefficient based on the meta-path, it is normalized by the softmax function. The normalized attention coefficient The calculation is as follows:

[0161]

[0162] Further, the node level coefficient matrix is ​​formed Node-level coefficient matrix Can be directly used with embedded tensors Multiplication:

[0163]

[0164] in, is the metapath φ p The embedding tensor learned above. The embedding of each node can be obtained by aggregating its neighbors.

[0165] For a given meta-path, the neighbors of each node play different roles in graph embedding for specific tasks and show different importance. Therefore, introducing node-level attention can learn to aggregate the importance of each node to its neighbors based on the meta-path.

[0166] The third step is to perform meta-path-level attention aggregation operations, including:

[0167] Determining the importance of meta paths Calculated using the following formula:

[0168]

[0169] Where Q is the learnable parameter matrix, b is the bias, and r is the meta-path-level attention vector.

[0170] After obtaining the importance of each meta-path, it is normalized by the softmax function. The normalized value of the meta-path level attention coefficient is expressed as:

[0171]

[0172] The learned weights are used as coefficients to merge these specific embeddings to obtain the final representation matrix of layer l

[0173]

[0174] In particular, the feature tensor of the first layer (i.e., the input layer) is equal to the normalized data matrix, i.e. The feature tensor of the last layer (i.e., the output layer) is equal to the output node embedding matrix Z, that is,

[0175] The node embedding matrix Z can be expanded as:

[0176] Z=[Z1;Z2;…;Z i ;…Z N ] (17)

[0177] Among them, Z i represents the embedding vector of node i, whose value is the i-th row of Z; Z N represents the embedding vector of node N, whose value is the Nth row of Z.

[0178] Step S404b: generating a reconstructed adjacency matrix according to the node embedding matrix and a preset reconstruction matrix generation algorithm.

[0179] In this embodiment, the preset reconstruction matrix generation algorithm includes calculating a cosine similarity matrix between any two nodes and generating a reconstruction adjacency matrix according to the cosine similarity matrix.

[0180] Specifically, based on the node embedding matrix Z, the cosine similarity between node i and node j can be calculated according to the following formula:

[0181]

[0182] The “2” in ∥∥2 represents the L2 norm.

[0183] The node similarity matrix C can be obtained by c i,j The data in row i and column j is equal to c i,j The value of .

[0184] All c i,j The values ​​of are arranged in descending order, and the kth largest value is taken as the threshold h, then the adjacency matrix is ​​reconstructed It can be obtained from the cosine similarity matrix C, if The element in row i and column j in If it is greater than the threshold h, it is set to 1, otherwise it is set to 0. The specific formula can be expressed as:

[0185]

[0186] Step S404c: Calculate the loss value of the spatiotemporal heterogeneous graph attention neural network model using the reconstructed adjacency matrix and the true adjacency matrix.

[0187] In one example, the value L of the loss function can be represented by the L2 norm of the cosine similarity matrix and the true adjacency matrix, and its calculation formula is as follows:

[0188]

[0189] Step S404d: Determine whether the loss value satisfies the preset loss condition. If so, obtain the target wireless data knowledge graph model; otherwise, continue training.

[0190] In one example, the preset loss condition may be that the loss function values ​​for 20 consecutive rounds are all less than the loss function threshold ∈. If this condition is met, training is stopped, and the reconstructed adjacency matrix at this time is output as the final predicted adjacency matrix. Otherwise, the process returns to step S204 to continue training.

[0191] In a specific embodiment, taking uplink throughput as an example, the specific process of constructing a wireless data knowledge graph includes the following steps:

[0192] Step 1: Build an uplink throughput knowledge graph model based on 82 data fields related to uplink throughput. There are 82 nodes and 133 edges, and each node corresponds to a data field.

[0193] In the second step, the real data matrix of uplink throughput in the uplink throughput knowledge graph model is interpolated and averaged to obtain the standardized real data matrix of uplink throughput.

[0194] Step 3: Randomly select 70% of the 82*82=6724 points in the adjacency matrix of the uplink throughput knowledge graph, or 4707 points, as the training set. The remaining 30%, or 2017 points, are used as the test set. This random selection process is repeated 30 times to obtain 30 training samples. At the same time, the corresponding normalized real data is used as the input process data.

[0195] The fourth step is to initialize the parameters of the spatiotemporal heterogeneous graph attention neural network model;

[0196] Step 5: Calculate the graph convolution kernel based on the graph structure of the uplink throughput knowledge graph model;

[0197] In the sixth step, the training set data and graph convolution kernel are input into the spatiotemporal heterogeneous graph attention neural network model to perform temporal convolution calculation and graph attention convolution calculation respectively to obtain the node embedding vector matrix;

[0198] The seventh step is to calculate the cosine similarity matrix of any two nodes based on the node embedding matrix.

[0199] Step 8: Arrange the elements in the cosine similarity matrix in descending order, set the elements greater than the threshold h to 1, and vice versa to 0 to obtain the reconstructed adjacency matrix;

[0200] In the ninth step, the L2 norm of the reconstructed adjacency matrix and the true adjacency matrix is ​​calculated as the loss value. If the loss function value for 20 consecutive rounds is less than the loss function threshold ∈, the training is stopped. Otherwise, return to the sixth step to continue training.

[0201] The embodiment of the present invention constructs a wireless data knowledge graph based on a small amount of data, and through data training of the spatiotemporal heterogeneous graph attention neural network model, analyzes and infers the relationship between each data, continuously enriches the wireless data knowledge graph until a complete wireless data knowledge graph is constructed, providing rich data support for subsequent AI models.

[0202] This embodiment of the present invention also includes: dividing the local wireless data knowledge graph model into multiple graph slice models according to coherence time.

[0203] In this embodiment, T = {t1, t2, ...} mainly reflects the time nature of the wireless data knowledge graph, where 0 <t1<t2<…,且t i ∈{t1,t2,…} is the sampling time. In addition, t i and t i+1 Indicates adjacent sampling times, t1, t2, and all subsequent t i The superposition forms a continuous sampling time period. The wireless data knowledge graph may have different graph structures at these sampling moments, that is, the wireless data knowledge graph is a continuous time dynamic graph.

[0204] The wireless data knowledge graph is modeled as a sequence of time-stamped events Represents the graph structure corresponding to each communication instance. The graph structure corresponding to each sampling time may be the same or different. According to the communication principle, within each coherence time, the channel state of the wireless communication network is stable. Therefore, within the coherence time T cWithin the time frame T, the topological structure of the wireless data knowledge graph can be considered to be stable. c , the wireless data knowledge graph can be divided into discrete graph slices, such as Figure 7 As shown. The coherence time switching point is mT c , where m is a natural number. In other words, from (m-1)T c to mT c -1 shares the same topological structure, determined by the above construction process. The mth graph slice is represented as The total number of graph slices is M. Therefore, the wireless data knowledge graph within the sampling period T can be modeled as a series of graph slices Graph Slices Represents many sampling instances, each instance corresponds to a G(t j ), indicating that these G(t j )’s graph structure is The inside remains unchanged.

[0205] In one example, the mth graph slice can be obtained using Indicates that and W remain unchanged; because in the wireless data knowledge graph, the number of nodes and node attributes remain consistent over time. A represents the adjacency matrix of the wireless data knowledge graph, and (A) i,j Represents the element in row i and column j of A. If there is an edge between node i and node j, then (A) i,j If yes, it is 1, otherwise it is 0. m Representing graph slices The adjacency matrix of X represents the matrix formed by the real wireless data collected by each node in the wireless data knowledge graph. m For graph slices The corresponding real data matrix can collect wireless big data at each time t in a coherent time period. Let node v i The real data collected at time t is Therefore, the data collected by all N nodes at time t can form a data vector Accordingly, X m It can be expressed as

[0206] Correspondingly, performing normalization processing on the real data matrix of each node to obtain the normalized real data matrix of each node includes:

[0207] The node real data matrix corresponding to each graph slice model is standardized to obtain the standardized real data matrix of each node.

[0208] By dividing the wireless data knowledge graph into graph slices according to coherent time, not only the characteristics of the wireless data knowledge graph are retained, but also simpler and more optimized data support is provided for subsequent data processing.

[0209] like Figure 8 As shown, the real-time extraction of key feature datasets in the target wireless data knowledge graph model according to the feature dataset generation algorithm includes the following steps:

[0210] Step S801: Obtain a set of influences of each node in the target wireless data knowledge graph model on the target KPI node.

[0211] Step S802: Sort the nodes in the influence set.

[0212] In this embodiment, the influence of other nodes in the wireless data knowledge graph model on the target KPI node can be calculated first, and then the influence of each node can be sorted in descending order to obtain an influence set.

[0213] In a specific embodiment, the first step is to determine the weight of each edge of the target wireless data knowledge graph model.

[0214] In one example, the similarity c obtained by similarity calculation formula (18) can be i,j As the edge weight between nodes i and j with an edge relationship.

[0215] The second step is to calculate the influence based on the edge weight.

[0216] In one example, for any (i, j)∈N×N, (i, j) represents a directed edge from node i to node j. According to the edge weight ω(i, j)=c ij (21). Similarly, ω(j,i)=c ji (22) represents the weight of the directed edge (j, i). At this time, the influence calculation process between any two points in the wireless data knowledge graph is as follows:

[0217] The degree of influence e of node i on node j (node ​​i is the mth-order neighbor node of node j) ij The formula is:

[0218]

[0219] in,

[0220] Represents the product of all edge weights on the t-th shortest path from node i to node j.

[0221] Therefore, for the target KPI node p, the influence degree set of all nodes on it can be obtained as {e 1p ,e 2p ,…,e np}.

[0222] Step S803: taking the node feature with the highest influence ranking as the dependent variable and inputting it into the KPI prediction algorithm to obtain the KPI prediction value;

[0223] Step S804: Calculate the degree of fit based on the KPI prediction value, the actual KPI value and the degree of fit calculation rule.

[0224] The goodness of fit d formula is as follows:

[0225]

[0226] Among them, Y represents the true value, Y' represents the predicted value, Represents the mean value of Y.

[0227] Step S805: comparing the fitness with a preset fitness;

[0228] Step S806: If the degree of fit is less than the preset degree of fit, continue to add the features of the next node in the influence set as a common dependent variable, and recalculate the KPI prediction value and the degree of fit until it is equal to or greater than the preset degree of fit.

[0229] In one example, the preset fitness is d0, if d <d0,则添加影响度排序中的下一个特征,并结合第一个特征再次预测KPI。当达到d0时,该过程停止,如果未达到预设拟合度d0,则继续添加特征,直到目标达成。

[0230] Step S807: Generate a key feature data set of the target business based on the nodes that have reached a preset degree of fit and the real data collected by the nodes.

[0231] In one example, if the predicted values ​​calculated by the first three nodes in the influence ranking reach a good fit, then these three nodes and the real data collected by the nodes are used as the key feature data set of the target business.

[0232] By setting the fit index, the most important features are selected as much as possible. These features, namely the relevant nodes in the graph, are extracted from these nodes and the real data collected by the nodes to form a feature dataset for the target KPI node.

[0233] In a specific embodiment, the method for generating a feature dataset includes the following steps:

[0234] In the first step, the cosine similarity between the target KPI node and other nodes is used as the edge weight between the two nodes;

[0235] The second step is to calculate the product of all edge weights on the shortest path between the target KPI node and any other node, and use this product as the influence of one of the nodes on the KPI target node; the influence of all nodes on the target KPI node is calculated using the same method.

[0236] The third step is to sort the obtained influences in descending order;

[0237] The fourth step is to input the node data ranked first into the KPI prediction algorithm to obtain the KPI prediction value;

[0238] Step 5: Calculate the fitting degree d between the actual KPI value and the predicted value according to the fitting disturbance formula;

[0239] The sixth step is to compare the fitness with the preset fitness d0. If d <d0,则添加影响度排序中的下一个特征,并结合第一个特征再次预测KPI。当达到d0时,该过程停止,如果未达到预设拟合度d0,则继续添加特征,直到目标达成。

[0240] In the seventh step, all node data participating in the dependent variable calculation that reaches the preset fitting degree are used as the key feature data set.

[0241] The present invention can extract a small amount of key data from massive and complex data through a feature data set generation method, and train and infer the AI ​​model through a small amount of key data, which greatly reduces the computing cost, saves computing power, and ensures the efficiency of real-time communication in the 6G network.

[0242] The embodiment of the present invention further includes: calculating a feature compression rate based on the number of extracted nodes that achieve a fitting degree and the total number of nodes in the wireless data knowledge graph;

[0243] The key feature dataset is evaluated using the feature compression ratio.

[0244] In this embodiment, the feature compression rate refers to extracting k features from the original feature N, that is, achieving an effect of a fitting degree greater than d, and its calculation formula is (Nk) / N.

[0245] The knowledge graph generation module also includes:

[0246] A construction unit, configured to construct a local wireless data knowledge graph model; the local wireless data knowledge graph model at least includes a real adjacency matrix and a real data matrix of each node;

[0247] A processing unit, configured to perform standardization processing on the real data matrix of each node to obtain a standardized real data matrix of each node;

[0248] A partitioning unit, configured to generate a training set from the normalized real data matrix of each node;

[0249] The training unit is used to input the training set into a preset neural network model for training to obtain a target wireless data knowledge graph model.

[0250] The training unit is further used to input the training set into a preset neural network model for training to obtain a node embedding matrix;

[0251] Generate a reconstructed adjacency matrix according to the node embedding matrix and a preset reconstruction matrix generation algorithm;

[0252] Calculating the loss value of the spatiotemporal heterogeneous graph attention neural network model using the reconstructed adjacency matrix and the true adjacency matrix;

[0253] Determine whether the loss value meets the preset loss condition. If so, obtain the target wireless data knowledge graph model; otherwise, continue training.

[0254] In an embodiment of the present invention, the wireless knowledge graph construction device also includes a slicing unit for dividing the local wireless data knowledge graph model into multiple graph slice models according to coherence time.

[0255] The feature dataset generation module also includes:

[0256] The prediction unit is used to use the node feature with the highest influence as the dependent variable and input it into the KPI prediction algorithm to obtain the KPI prediction value;

[0257] A fitting unit, configured to calculate a degree of fit based on the KPI predicted value, the actual KPI value, and a degree of fit calculation rule;

[0258] a comparing unit, configured to compare the degree of fit with a preset degree of fit;

[0259] A second calculation unit is configured to, when the degree of fit is less than the preset degree of fit, continue to add the feature of the next node in the influence set as a common dependent variable, and recalculate the KPI prediction value and the degree of fit until the KPI prediction value and the degree of fit are equal to or greater than the preset degree of fit;

[0260] The second generating unit is configured to generate a key feature data set of a target business based on the nodes that have reached a preset degree of fit and the real data collected by the nodes.

[0261] In an embodiment of the present invention, the feature data set generation device based on the wireless data knowledge graph also includes a feature evaluation unit, which is used to calculate the feature compression rate based on the number of extracted nodes that reach the fitting degree and the total number of nodes in the wireless data knowledge graph; and use the feature compression rate to evaluate the key feature data set.

[0262] Figure 9 An example of a physical structure diagram of an electronic device is shown below. Figure 9 As shown, the electronic device may include: a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other via the communication bus 940. The processor 910 may call the logic instructions in the memory 930 to execute the steps of any of the above methods.

[0263] In addition, the logic instructions in the above-mentioned memory 930 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0264] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the specific steps provided by the above methods.

[0265] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is used to execute the specific steps provided in the above methods when the computer program is executed by a processor.

[0266] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0267] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0268] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An endogenous intelligent system based on wireless data knowledge graph, characterized by: include: Wireless data knowledge graph intelligent generation unit, feature data set generation unit, wireless network digital twin unit and AI model unit; The wireless data knowledge graph intelligent generation unit is used to input local wireless communication data in the business scenario into a preset neural network model to generate a target wireless data knowledge graph model; The feature data set generation unit is used to extract the key feature data set of the target service in real time from the target wireless data knowledge graph according to the feature data set generation algorithm; The key feature dataset is the minimum valid dataset associated with the target business; The AI ​​model unit is used to generate an initial target service policy based on the key feature data set, and send the initial target service policy to the wireless network digital twin unit for policy pre-verification; The wireless network digital twin unit is used to perform policy pre-verification based on the key feature data set and the initial target service policy, and feed back the pre-verification result to the AI ​​model unit, so that the AI ​​model unit adjusts the final target service policy based on the pre-verification feedback result; The feature data set generation unit includes a feature sorting module and a feature data set generation module; The feature sorting module is used to perform feature sorting based on the influence of other nodes in the target wireless data knowledge graph model on the target KPI node; The feature dataset generation module also includes: The prediction unit is used to use the node feature with the highest influence as the dependent variable and input it into the KPI prediction algorithm to obtain the KPI prediction value; A fitting unit, configured to calculate a degree of fit based on the KPI predicted value, the actual KPI value, and a degree of fit calculation rule; a comparing unit, configured to compare the degree of fit with a preset degree of fit; A second calculation unit is configured to, when the degree of fit is less than the preset degree of fit, continue to add the feature of the next node in the influence set as a common dependent variable, and recalculate the KPI prediction value and the degree of fit until the KPI prediction value and the degree of fit are equal to or greater than the preset degree of fit; The second generating unit is configured to generate a key feature data set of a target business based on the nodes that have reached a preset degree of fit and the real data collected by the nodes.

2. The endogenous intelligent system based on wireless data knowledge graph according to claim 1 is characterized in that: The wireless data knowledge graph intelligent generation unit includes a non-real-time data acquisition module, a data preprocessing module and a knowledge graph generation module; The non-real-time data acquisition module is used to collect multiple types of non-real-time data in business scenarios through a combination of one or more of hard acquisition, soft acquisition or drive testing; The data preprocessing module performs one or more operations of deleting, distinguishing by time, and structurally processing the multiple types of non-real-time data to obtain preprocessed data; The knowledge graph generation module is used to construct a local wireless data knowledge graph model based on the preprocessed data, and generate a target wireless data knowledge graph model based on the local wireless data knowledge graph model and a preset neural network model.

3. The endogenous intelligent system based on wireless data knowledge graph according to claim 1 is characterized in that: The feature data set generating unit includes a feature data set evaluating module, The feature data set evaluation module is used to evaluate the key feature data set.

4. A method for constructing an endogenous intelligent system based on wireless data knowledge graph, characterized in that: include: Generate a target wireless data knowledge graph model based on local wireless communication data in the business scenario and a preset neural network model; According to the feature data set generation algorithm, the key feature data set of the target business is extracted in real time from the target wireless data knowledge graph model; The key feature dataset is the minimum valid dataset associated with the target business; Inputting the key feature dataset into the AI ​​model to generate an initial target business strategy; Inputting the initial target service strategy and the key feature data set into the wireless network digital twin for pre-verification; Feeding pre-verification results back to the AI ​​model; Adjust the initial target business strategy based on the pre-verification feedback results until the final target business strategy is issued; Extracting the key feature dataset of the target service from the wireless data knowledge graph according to the feature dataset generation algorithm includes: Obtain the influence degree set of each node on the target KPI node in the target wireless data knowledge graph model; Sorting the nodes in the influence set; The node feature with the highest influence ranking is used as the dependent variable and input into the KPI prediction algorithm to obtain the KPI prediction value; Calculate the degree of fit based on the KPI predicted value, the actual KPI value and the degree of fit calculation rule; comparing the fitness with a preset fitness; If the degree of fit is less than the preset degree of fit, then continue to add the feature of the next node in the influence set as a common dependent variable, and recalculate the KPI prediction value and the degree of fit until it is equal to or greater than the preset degree of fit; Generate key feature data sets for the target business based on the nodes that have achieved the preset fit and the real data collected by the nodes.

5. The method for constructing an endogenous intelligent system based on wireless data knowledge graph according to claim 4 is characterized in that: The target wireless data knowledge graph model is generated based on the local wireless communication data in the business scenario and the preset neural network model, including: Constructing a local wireless data knowledge graph model; the local wireless data knowledge graph model at least includes a real adjacency matrix and a real data matrix of each node; Performing standardization processing on the real data matrix of each node to obtain a standardized real data matrix of each node; Generate a training set based on the standardized real data matrix of each node; The training set is input into a preset neural network model for training to obtain a target wireless data knowledge graph model.

6. The method for constructing an endogenous intelligent system based on wireless data knowledge graph according to claim 5 is characterized in that: Inputting the training set into a preset neural network model for training to obtain a target wireless data knowledge graph model includes: Inputting the training set into a preset neural network model for training to obtain a node embedding matrix; Generate a reconstructed adjacency matrix according to the node embedding matrix and a preset reconstruction matrix generation algorithm; Calculating the loss value of the spatiotemporal heterogeneous graph attention neural network model using the reconstructed adjacency matrix and the true adjacency matrix; Determine whether the loss value meets the preset loss condition. If so, obtain the target wireless data knowledge graph model; otherwise, continue training.

7. The method for constructing an endogenous intelligent system based on wireless data knowledge graph according to claim 4 is characterized in that: The method further comprises: Dividing the local wireless data knowledge graph model into multiple graph slice models according to coherence time; Correspondingly, performing normalization processing on the real data matrix of each node to obtain the normalized real data matrix of each node includes: The node real data matrix corresponding to each graph slice model is standardized to obtain the standardized real data matrix of each node.

8. The method for constructing an endogenous intelligent system based on wireless data knowledge graph according to claim 4 is characterized in that: The method further comprises: The feature compression rate is calculated based on the number of extracted nodes that reach the fitting degree and the total number of nodes in the wireless data knowledge graph; The key feature dataset is evaluated using the feature compression ratio.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 4 to 8 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 4 to 8 is implemented.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 4 to 8 is implemented.

Images