Wireless network load dynamic adjustment method, training method, device and electronic equipment

Abstract

The invention provides a wireless network load dynamic adjustment method and device, a training method and device and electronic equipment. The adjustment method comprises the following steps: acquiring model parameters of a machine learning model of at least one cell of a centralized computing node sending guarantee area; loading the model parameters to obtain a machine learning model of the corresponding cell; processing the telephone traffic load data of the corresponding cell in real time by using a machine learning model to obtain a telephone traffic decision point feature value; converting the telephone traffic decision point characteristic value into a base station parameter adjustment command; sending the base station parameter adjustment command to the base station, so that the base station adjusts a corresponding base station parameter according to the base station parameter adjustment command; evaluating the real-time data obtained after the base station parameters are adjusted, and determining a threshold adjustment instruction; and sending a threshold adjustment instruction to the centralized computing node to enable the centralized computing node to perform corresponding threshold adjustment, retraining the machine learning model, and sending model parameters of the retrained machine learning model to the edge computing node.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a wireless network load dynamic adjustment method, a model training method, a device and an electronic device. Background Art

[0002] With the rapid development of wireless communication technology, wireless networks are becoming more mature, and network planning and large-scale investment and construction are basically completed. However, the growing user base, network services, and the imbalance of user and service distribution make the early planning increasingly unable to meet the actual needs of wireless networks.

[0003] In the related technology, for wireless network planning, machine learning algorithms are usually used to predict the trend of network load changes in the future based on historical traffic data. Based on the prediction results, wireless network adjustments are made by adding cells or expanding capacity. However, these adjustment methods are time-consuming and have poor real-time performance, making it difficult to adapt to high-real-time scenario requirements such as emergency support.

[0004] Therefore, based on the high real-time scenario requirements, how to dynamically adjust the wireless network load is a problem that needs to be solved. Summary of the invention

[0005] The present invention provides a method for dynamically adjusting wireless network load, a model training method, a device and an electronic device, so as to at least solve the problem that in the related art, for scenarios with high real-time requirements, wireless network adjustment can only be performed by adding cells or expanding capacity, resulting in a long adjustment time and poor real-time performance. The technical solution of the present invention is as follows:

[0006] According to a first aspect of an embodiment of the present invention, a method for dynamically adjusting a wireless network load is provided, the method being applied to an edge computing node, comprising:

[0007] Obtain model parameters of a machine learning model of at least one cell in a security area sent by a centralized computing node;

[0008] Load the model parameters to obtain a machine learning model for the corresponding cell, where the machine learning model is a model trained based on a multi-layer perception MLP neural network;

[0009] Using the machine learning model to process the traffic load data of the corresponding cell in real time to obtain a traffic decision point characteristic value;

[0010] Converting the traffic decision point characteristic value into a base station parameter adjustment command;

[0011] Sending the base station parameter adjustment command to a base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command;

[0012] Evaluating the real-time data obtained after adjusting the base station parameters to determine a threshold adjustment instruction;

[0013] The threshold adjustment instruction is sent to the centralized computing node, so that the centralized computing node performs corresponding threshold adjustment based on the threshold adjustment instruction, retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node.

[0014] According to a second aspect of an embodiment of the present invention, a method for training a machine learning model is provided, the method being applied to a centralized computing node, comprising:

[0015] Initialize the model parameters of the machine learning model;

[0016] Get training data set and test data set;

[0017] Based on the training data set, the machine learning model after initializing the model parameters is trained for multiple rounds, wherein each round of training includes: initializing the measurement variables of the training data set, reading the training data set and its label, and storing the label value of the label; traversing the training data set, during the traversal process, calculating the characteristic value using the model parameters, and storing the calculated characteristic value as the characteristic value output of this batch of training data, using the loss function to calculate the loss value, calculating the gradient based on the loss value, performing back propagation, and updating the machine learning model function, updating the measurement variables, and storing the loss value as the calculation variable; reading the stored label value and the characteristic value output of this batch of training data, calculating the accurate number of this batch of training data based on the label value and the characteristic value, and updating the accurate number and the total number to the measurement variable, and superimposing the original value to obtain the calculation variable including the accurate number, the total number and the loss value; repeating the above process until all the theoretical training is completed, and outputting the accuracy of the corresponding training result;

[0018] When the accuracy reaches a threshold, the machine learning model is determined to be a trained model.

[0019] According to a third aspect of an embodiment of the present invention, a wireless network load dynamic adjustment device is provided, and the device is applied to an edge computing node, including:

[0020] A parameter acquisition module, used to acquire model parameters of a machine learning model of at least one cell in a security area sent by a centralized computing node;

[0021] A loading module, used to load the model parameters to obtain a machine learning model of the corresponding cell, wherein the machine learning model is a model obtained by training based on a multi-layer perception MLP neural network;

[0022] A processing module, used to use the machine learning model to process the traffic load data of the corresponding cell in real time to obtain a traffic decision point characteristic value;

[0023] A conversion module, used for converting the traffic decision point characteristic value into a base station parameter adjustment command;

[0024] A first sending module, configured to send the base station parameter adjustment command to a base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command;

[0025] An evaluation module, used to evaluate the real-time data obtained after adjusting the base station parameters and determine a threshold adjustment instruction;

[0026] The second sending module is used to send the threshold adjustment instruction to the centralized computing node, so that the centralized computing node performs corresponding threshold adjustment based on the threshold adjustment instruction, retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node.

[0027] According to a fourth aspect of an embodiment of the present invention, a training device for a machine learning model is provided, the device being applied to a centralized computing node, comprising:

[0028] Initialization module, used to initialize the model parameters of the machine learning model;

[0029] An acquisition module is used to obtain training data sets and test data sets;

[0030] A training module is used to perform multiple rounds of training on the machine learning model after initializing the model parameters based on the training data set; wherein each round of training includes: initializing the measurement variables of the training data set, reading the training data set and its labels, and storing the label values ​​of the labels; traversing the training data set, during the traversal process, calculating the characteristic values ​​using the model parameters, and storing the calculated characteristic values ​​as the characteristic values ​​output by this batch of training data, using the loss function to calculate the loss value, calculating the gradient based on the loss value, performing back propagation, and updating the machine learning model function, updating the measurement variables, and storing the loss value as a calculation variable; reading the stored label values ​​and the characteristic values ​​output by this batch of training data, calculating the accurate number of this batch of training data based on the label values ​​and the characteristic values, and updating the accurate number and the total number into the measurement variables, and superimposing the original values ​​to obtain a calculation variable including the accurate number, the total number and the loss value; repeating the above process until all the theoretical training is completed, and outputting the accuracy of the corresponding training results;

[0031] A determination module is used to determine that the machine learning model is a trained model when the accuracy reaches a threshold.

[0032] According to a fifth aspect of an embodiment of the present invention, there is provided an electronic device, including:

[0033] processor;

[0034] a memory for storing instructions executable by the processor;

[0035] Wherein, the processor is configured to execute the instructions to implement the wireless network load dynamic adjustment method as described above or the machine learning model training method as described above.

[0036] According to the sixth aspect of an embodiment of the present invention, a computer-readable storage medium is provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the wireless network load dynamic adjustment method as described above or the machine learning model training method as described above.

[0037] According to a seventh aspect of an embodiment of the present invention, a computer program product is provided, comprising a computer program or instructions, which, when executed by a processor of an electronic device, implements the method for dynamically adjusting the wireless network load as described above or the method for training the machine learning model as described above.

[0038] The technical solution provided by the embodiments of the present invention brings at least the following beneficial effects:

[0039] In an embodiment of the present invention, the model parameters of the machine learning model of at least one cell in the protection area sent by the centralized computing node are obtained; the model parameters are loaded to obtain the machine learning model of the corresponding cell, and the machine learning model is a model obtained by training based on the multi-layer perception MLP neural network; the machine learning model is used to process the traffic load data of the corresponding cell in real time to obtain the characteristic value of the traffic decision point; the characteristic value of the traffic decision point is converted into a base station parameter adjustment command; the base station parameter adjustment command is sent to the base station so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command; the real-time data obtained after adjusting the base station parameters is evaluated to determine the threshold adjustment instruction; the threshold adjustment instruction is sent to the centralized computing node so that the centralized computing node performs the corresponding threshold adjustment based on the threshold adjustment instruction, and retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node. That is to say, the embodiment of the present invention uses the machine learning model to convert the predicted object from the wireless network load indicator itself into the "feature" of the wireless network load indicator, and converts the "feature" into the base station parameter adjustment command, thereby realizing the dynamic adjustment of the wireless network load. Since this conversion has strong real-time performance, it can adapt to scenarios with high real-time requirements such as emergency protection. The embodiment of the present invention also splits the mapping relationship between the base station index and the optimization adjustment strategy into the relationship between the base station index and the "feature", and the mapping relationship between the "feature" and the optimization adjustment strategy. The implementation method is to adjust the parameter threshold of the marking algorithm, and then change the relationship between the base station index and the "feature" by re-learning the model during the guarantee period. In this way, the mapping relationship between the base station index and the optimization adjustment strategy becomes dynamic, which can adapt to the dynamic changes of the wireless network, making the effect of dynamic adjustment of the wireless network more accurate. The present invention also defines the signaling interaction format between the centralized computing node, the edge computing node and the BBU, as well as the format and update method of various databases for storing key data information of each network element on the basis of establishing a federated learning technology architecture composed of centralized computing nodes and edge computing nodes, which can effectively handle instruction execution failures and dynamic changes in various network parameters, and ensure the stable operation of the entire technical system.

[0040] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention, and do not constitute an improper limitation of the present invention. In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 The present invention provides a flowchart of a method for dynamically adjusting wireless network load.

[0043] Figure 2 It is a flow chart of a model training method provided by an embodiment of the present invention.

[0044] Figure 3 It is an application flow chart of a method for dynamically adjusting wireless network load provided by an embodiment of the present invention;

[0045] Figure 4 It is a schematic diagram of a computing architecture established in combination with a 5G wireless network architecture provided by an embodiment of the present invention.

[0046] FIG5(1) is a schematic diagram of a hierarchical structure of an MLP model adopted in an embodiment of the present invention.

[0047] FIG5(2) is a processing flow chart of an MLP model adopted in an embodiment of the present invention.

[0048] Figure 6 The present invention provides a flowchart of writing a single indicator file into a training data set or a test data set.

[0049] Figure 7 It is a schematic diagram of a data reorganization and labeling process provided by an embodiment of the present invention.

[0050] Figure 8 It is an overall algorithm flow chart of MLP model parameter training provided by an embodiment of the present invention.

[0051] Figure 9-1 The embodiment of the present invention provides Figure 8 The algorithm flow of performing a training on the training data set during the parameter training process.

[0052] Figure 9-2 The embodiment of the present invention provides Figure 9-1 The algorithm flow for calculating the accuracy of this batch of training data.

[0053] Fig.10 It is an algorithm flow for calculating the accuracy of test data during the MLP model parameter training process provided by an embodiment of the present invention.

[0054] Fig.11 This embodiment of the present invention provides a Fig.10 The algorithm flow for calculating the accuracy of this batch of test data.

[0055] Fig.12 It is the initial training loss rate and accuracy change curve of Example 1 provided in the embodiment of the present invention.

[0056] Fig.13 It is the initial training loss rate and accuracy change curve of Example 2 provided in the embodiment of the present invention.

[0057] Fig.14 It is a signaling interaction diagram for defining a model parameter sending process provided by an embodiment of the present invention.

[0058] Fig.15 It is an instruction execution signaling interaction diagram between an edge computing node and a BBU provided by an embodiment of the present invention.

[0059] Fig.16 It is a signaling interaction diagram for updating marking threshold parameters and adding test data in model training during a guarantee period provided by an embodiment of the present invention.

[0060] Fig.17 The invention is a block diagram of a device for dynamically adjusting wireless network load provided by an embodiment of the invention.

[0061] Fig.18 It is a block diagram of a model training device provided by an embodiment of the present invention.

[0062] Fig.19 It is a block diagram of an electronic device provided by an embodiment of the present invention.

[0063] Fig. 20 It is a block diagram of a device for dynamic adjustment of wireless network load or model training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable ordinary persons in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0065] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0066] Technical terms:

[0067] PRB (Physical Resource Block): In 5G, PRB is a combination of 12 consecutive subcarriers in the frequency domain and a slot in the time domain (14 OFDM symbols in the case of conventional CP, but 12 OFDM symbols in the case of extended CP). It is the smallest unit for data transmission and resource allocation at the physical layer.

[0068] 5G PRB utilization: refers to the proportion of physical radio resource blocks (PRBs) that are effectively utilized within a certain time window in a 5G network. The calculation formula is: total number of occupied PRBs / total number of available PRBs × 100%. Whether it is TDD or FDD, the allocation and use of uplink and downlink PRBs are separate, so it can be divided into uplink PRB utilization and downlink PRB utilization.

[0069] PDSCH (Physical Downlink Shared Channel): PDSCH is a type of physical downlink channel in wireless communication systems, specifically used to transmit downlink user data. It is part of the physical layer downlink shared layer, carrying user-specific data, and is a key channel for user data transmission in 5G networks.

[0070] PDCCH (Physical Downlink Control Channel): PDCCH is a key physical channel used to transmit downlink control information in the 5GNR (New Radio) system. It carries control information sent by the base station (gNodeB) to the end user equipment (UE), including scheduling allocation for PDSCH (Physical Downlink Shared Channel) reception, scheduling authorization for PUSCH (Physical Uplink Shared Channel) reception, as well as power control, time slot format indication, resource preemption indication and other information.

[0071] 5G Throughput: refers to the amount of data successfully transmitted from the sender to the receiver per unit time in a 5G network. It reflects the data transmission capacity and efficiency of the network and is one of the important indicators for evaluating the performance of 5G networks. The throughput is usually calculated by dividing the amount of data successfully transmitted per unit time (in bits or bytes) by the length of time (such as seconds). Like PRB utilization, throughput is also divided into uplink and downlink, and can be counted for PDCP, RLC, MAC, physical layer and other layers. For example, for the MAC layer, the uplink MAC layer throughput and the downlink MAC layer throughput can be counted.

[0072] CCE (Control Channel Element): CCE is a control channel logical unit consisting of a group of resource element groups (REG). Each CCE contains a fixed number of REGs, which are mapped to specific physical resources in the frequency domain and time domain to form a PDCCH channel, and then transmit DCI to the end user equipment (UE).

[0073] MAC (Medium Access Control) layer: In 5G networks, the MAC (Medium Access Control) layer, as a key component of the wireless protocol stack, is responsible for mapping between logical channels and transmission channels, data multiplexing and demultiplexing, wireless resource allocation and management, scheduling information reporting, HARQ error correction, logical channel priority management and other functions.

[0074] Tensor: In machine learning, tensor is a core concept. It can be regarded as a generalized form of multidimensional arrays, which is used to represent and manipulate complex data structures. Tensor is the core structure used to represent data in deep learning and machine learning. It can be regarded as a generalization of multidimensional arrays. This means that tensors are not limited to two dimensions (matrices) or one dimension (vectors), but can be extended to higher dimensions. In machine learning models, tensors are used to store and transform data and are the cornerstone of implementing complex algorithms. They are able to represent and store many types of data including images, videos, audio, etc.

[0075] Model: In machine learning, model is a core concept, which refers to a mathematical function or a set of mathematical functions that can map input data to predicted output. Specifically, the model is the core part of the machine learning algorithm, which adaptively adjusts the model parameters by learning training data to minimize the error between the predicted output and the true label.

[0076] Neural Network model: It imitates the structure of biological nervous system and establishes complex mapping relationships through connections between multiple layers of neurons, such as multi-layer perceptron, convolutional neural network (CNN), recurrent neural network (RNN), etc.

[0077] Label: In machine learning, a label is a value that marks a specific attribute or category of a data sample. It is a key component of the supervised learning process, used to guide the model to learn how to predict the corresponding output from the input data. Labels can be discrete (such as categories in classification problems), continuous (such as numerical values ​​in regression problems), or multi-valued (such as a set of labels in multi-label classification problems).

[0078] Feature: A feature can be considered as a representation of raw data, which is preprocessed and converted into a format suitable for machine learning algorithms. Features can be numerical (such as age, height, price, etc.) or categorical (such as gender, color, country, etc.). For categorical features, encoding processing (such as one-hot encoding, label encoding, etc.) is usually required so that machine learning algorithms can process them.

[0079] Multilayer Perceptron (MLP): A deep learning model based on a feedforward neural network. A multilayer perceptron consists of multiple neuron layers, each of which is fully connected to the previous layer.

[0080] Input layer: receives raw data features as input, and each input neuron corresponds to one feature.

[0081] Hidden layer: located between the input layer and the output layer, used to extract features and perform nonlinear transformations. A multilayer perceptron can contain one or more hidden layers, each of which consists of multiple neurons that process and transmit information through weights and activation functions.

[0082] Output layer: Usually used to produce the final output of the network. The number of neurons in the output layer depends on the nature of the task. For example, for binary classification problems, there is usually one neuron, and for multi-class classification problems, there are usually multiple neurons.

[0083] Forward propagation: Input data is fed into the input layer, then passes through each hidden layer in turn, and finally generates the output of the output layer. Each neuron in each layer calculates the output of its weighted input and nonlinear activation function.

[0084] Backpropagation: During the training process, the multilayer perceptron updates the weights and biases between neurons through the backpropagation algorithm. Specifically, the output of the output layer is compared with the expected output to generate an error value. This error is then backpropagated to the network, and the weights and biases of each layer are adjusted accordingly based on the error signal to minimize the error between the predicted output and the actual output.

[0085] On the basis of understanding the above technical terms, please refer to the following embodiments.

[0086] Figure 1 is a flow chart of a method for dynamically adjusting wireless network load provided by an embodiment of the present invention, wherein the method is applied to edge computing nodes, such as Figure 1 As shown, the method comprises the following steps:

[0087] Step 101: Obtain model parameters of a machine learning model of at least one cell in a protection area sent by a centralized computing node.

[0088] Step 102: Load the model parameters to obtain a machine learning model of the corresponding cell, where the machine learning model is a model trained based on a multi-layer perception MLP neural network.

[0089] Step 103: Use the machine learning model to process the traffic load data of the corresponding cell in real time to obtain the characteristic value of the traffic decision point.

[0090] Step 104: Convert the traffic decision point characteristic value into a base station parameter adjustment command.

[0091] Step 105: Send the base station parameter adjustment command to the base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command.

[0092] Step 106: Evaluate the real-time data obtained after adjusting the base station parameters and determine a threshold adjustment instruction.

[0093] Step 107: Send the threshold adjustment instruction to the centralized computing node, so that the centralized computing node performs threshold adjustment based on the threshold adjustment instruction and retrains the machine learning model.

[0094] The wireless network load dynamic adjustment method described in the present invention can be applied to terminals, servers, etc., without limitation herein. The terminal implementation device can be an electronic device such as a smart phone, a laptop computer, a tablet computer, a desktop computer, a personal digital assistant (PDA, Personal Digital Assistant) and a wearable device. The server can be an independent server or a server cluster, or a server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, intermediate services, domain name services, security services, content distribution networks, or a big data and artificial intelligence platform, etc., without limitation herein.

[0095] Combine the following Figure 1 , the specific implementation steps of a method for dynamically adjusting a wireless network load provided by an embodiment of the present invention are described in detail. This embodiment takes an edge computing node as an example. In practical applications, it can also be applied to other similar nodes, and this embodiment does not limit it.

[0096] In step 101, model parameters of a machine learning model of at least one cell in a protection area sent by a centralized computing node are obtained.

[0097] In this step, first, the edge computing node receives the model parameter request sent by the central computing node, and the model parameter request includes: a model number and a corresponding cell; of course, in actual applications, it is not limited to this.

[0098] Secondly, when the edge computing node finds the cell in the model parameter request in the cell list associated with the node itself, the edge computing node sends a confirmation message including the model number to the central computing node.

[0099] That is to say, the edge computing node queries the cell list under the BBU room where it is located based on the model parameter request; and when the cell in the model parameter request is found in the cell list, a confirmation message including the model number is sent to the centralized computing node; so that the centralized computing node performs a local query based on the model number to obtain the pth file of the model corresponding to the model number, wherein the pth file is in binary format; and sends the model parameters including the model number and the corresponding pth file to the edge computing node.

[0100] Again, the edge computing node receives the model parameters sent by the centralized computing node, where the model parameters include the model number and the corresponding pth file.

[0101] In this step, the message exchange mechanism between the centralized computing nodes and the edge computing nodes is implemented, which is conducive to the dynamic management of model parameters and supports the needs of multiple model parameter learning before and during assurance.

[0102] In step 102, the model parameters are loaded to obtain a machine learning model of the corresponding cell, wherein the machine learning model is a model obtained by training based on a multi-layer perception MLP neural network.

[0103] In this step, the edge computing node maintains the model parameters, loads the pth file corresponding to the model number, obtains the machine learning model of the corresponding cell, and stores the pth file of the model parameters.

[0104] In step 103, the machine learning model is used to process the traffic load data of the corresponding cell in real time to obtain the characteristic value of the traffic decision point.

[0105] Among them, in this step, first, the edge computing node obtains the traffic load data of the corresponding cell, and the traffic load data includes: total throughput, number of users and wireless resource usage, etc.; secondly, the total throughput, number of users and wireless resource usage are merged according to the collection time period of the corresponding cell; then, the tensor of the data points in the corresponding cell collection time period after merging is determined, such as 1×120×7, etc.; finally, the model parameters of the corresponding cell are used to calculate the tensor to obtain the characteristic value of the traffic decision point. The specific implementation process is detailed as follows.

[0106] In step 104, the TDP characteristic value is converted into a base station parameter adjustment command.

[0107] In this step, the edge computing node determines the value range of the uplink and downlink modes of the corresponding cell respectively by using the sliding window to reorganize and mark the reverse method based on the characteristic value of the traffic decision point; for example, the uplink mode C upllnk The value range is 0 to 3, and the linear mode C downllnk The value range of is 0 to 6, etc., but in specific applications, the range can be adaptively expanded or reduced based on the actual application situation, etc., and this embodiment does not limit it.

[0108] Afterwards, the edge computing node sets the value of the network parameter according to the corresponding network parameter setting rule based on the determined value range of the uplink and downlink modes of the corresponding cell. The specific setting process is detailed in the following embodiment.

[0109] Then, the edge computing node obtains the values ​​of the uplink and downlink modes based on which the command was last issued in the command execution success record table of the corresponding cell; compares the values ​​of the uplink and downlink modes based on which the command was last issued with the values ​​of the determined uplink and downlink modes, and generates the parameter modification command Q of the corresponding cell according to the comparison result. t; Specifically including: the edge computing node compares the value of the sum of the uplink and downlink modes with the determined value of the sum of the uplink and downlink modes; according to the comparison result, query the command execution success record table, assign the index of the key-value pair of all column parameter names that meet the conditions, and generate the parameter modification instruction Q of the corresponding cell t .

[0110] Finally, the parameter modification instruction of the corresponding cell is converted into a base station parameter adjustment command that can be recognized by the BBU in the base station. Specifically including:

[0111] The edge computing node generates a corresponding single-cell and single-parameter modification script based on each key-value pair of the parameter modification instruction of the corresponding cell; according to the different data types of the parameters, based on each key-value pair of the corresponding cell, the mapping relationship between the parameters and the network management modification instruction parameters is searched to obtain the corresponding command name, the first-level switch type IE and / or the second-level switch type IE; the corresponding command name, the first-level switch type IE and / or the second-level switch type IE are converted into a single parameter modification script of the corresponding cell, and the single parameter modification script is used as the base station parameter adjustment command.

[0112] Among them, the converting of the parameter modification instruction into a base station parameter adjustment command recognizable by the BBU in the base station also includes: merging the single parameter modification script according to the structure of the base station network management instruction to obtain a single-cell, multi-parameter modification script, wherein the two merged single parameter modification scripts belong to the same base station, the same BBU, and use the same command name.

[0113] In step 105, the base station parameter adjustment command is sent to a base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command.

[0114] Among them, in this step, the edge computing node sends the base station parameter adjustment command to the base station, so that the base station adjusts the corresponding base station parameters according to the received base station parameter adjustment command. The specific adjustment process is detailed as follows.

[0115] In step 106, the real-time data obtained after adjusting the base station parameters is evaluated to determine a threshold adjustment instruction.

[0116] In this step, the edge computing node obtains the real-time data after adjusting the base station parameters, and determines the average uplink and downlink rate of each household in each cell according to the set time; divides the set time into multiple time periods, and calculates the value of the uplink and downlink mode of the corresponding cell in each time period; determines the value of the uplink and downlink mode of each cell and the number of occurrences of the uplink and downlink modes according to each time period; merges the uplink and downlink modes of each cell and the corresponding number of occurrences; based on the merged uplink and downlink modes of each cell and the corresponding number of occurrences, queries the average uplink and downlink rate of each household in the corresponding cell; combines the average uplink and downlink rate of each household in all cells to obtain a corresponding array of average uplink and downlink rates for each household; compares the value of each element in the array of average uplink and downlink rates for each household with the corresponding target value, and generates an adjustment instruction for the corresponding threshold, for example, generates and T uplink，prb Adjustment instructions for three thresholds, etc.

[0117] In step 107, the threshold adjustment instruction is sent to the centralized computing node, so that the centralized computing node performs threshold adjustment based on the threshold adjustment instruction and retrains the machine learning model.

[0118] In this step, the edge computing node sends the threshold adjustment instruction to the centralized computing node, so that the centralized computing node performs a corresponding threshold adjustment based on the received threshold adjustment instruction, and retrains the machine learning model based on the corresponding threshold adjustment, and sends the model parameters of the trained machine learning model to the edge computing node again.

[0119] In an embodiment of the present invention, the model parameters of the machine learning model of at least one cell in the protection area sent by the centralized computing node are obtained; the model parameters are loaded to obtain the machine learning model of the corresponding cell, and the machine learning model is a model obtained by training based on the multi-layer perception MLP neural network; the machine learning model is used to process the traffic load data of the corresponding cell in real time to obtain the characteristic value of the traffic decision point; the characteristic value of the traffic decision point is converted into a base station parameter adjustment command; the base station parameter adjustment command is sent to the base station so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command; the real-time data obtained after adjusting the base station parameters is evaluated to determine the threshold adjustment instruction; the threshold adjustment instruction is sent to the centralized computing node so that the centralized computing node performs the corresponding threshold adjustment based on the threshold adjustment instruction, and retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node. That is to say, the embodiment of the present invention uses the machine learning model to convert the predicted object from the wireless network load indicator itself into the "feature" of the wireless network load indicator, and converts the "feature" into the base station parameter adjustment command, thereby realizing the dynamic adjustment of the wireless network load. Since this conversion has strong real-time performance, it can adapt to scenarios with high real-time requirements such as emergency protection. The embodiment of the present invention also splits the mapping relationship between the base station index and the optimization adjustment strategy into the relationship between the base station index and the "feature", and the mapping relationship between the "feature" and the optimization adjustment strategy. The implementation method is to adjust the parameter threshold of the marking algorithm, and then change the relationship between the base station index and the "feature" by re-learning the model during the guarantee period. In this way, the mapping relationship between the base station index and the optimization adjustment strategy becomes dynamic, which can adapt to the dynamic changes of the wireless network, making the effect of dynamic adjustment of the wireless network more accurate. The present invention also defines the signaling interaction format between the centralized computing node, the edge computing node and the BBU, as well as the format and update method of various databases for storing key data information of each network element on the basis of establishing a federated learning technology architecture composed of centralized computing nodes and edge computing nodes, which can effectively handle instruction execution failures and dynamic changes in various network parameters, and ensure the stable operation of the entire technical system.

[0120] Optionally, in another embodiment, based on the above embodiment, after sending the base station parameter adjustment command to the base station, the method may also include: obtaining the current control channel unit CCE utilization index of each cell; determining the current CCE load of the corresponding cell based on the CCE utilization index; when the current CCE load of the cell is greater than a set threshold, generating a channel occupancy adjustment instruction; sending the channel occupancy adjustment instruction to the base station, so that the base station adjusts the channel parameters based on the channel occupancy adjustment instruction.

[0121] Among them, determining the current CCE load situation of the corresponding cell based on the CCE utilization index includes: determining the total CCE utilization of each cell at each moment based on the CCE utilization index; determining the average CCE utilization of each cell within a set time period based on the total CCE utilization of each cell, and using the average CCE utilization as the current CCE load situation of the corresponding cell.

[0122] In this step, before the guarantee starts, as the default value, set the cell's "PDCCH channel occupied symbols" to 2 symbols, set the "PDCCH channel occupied RB number" to 0 (i.e. occupying the full bandwidth), and turn on the "PDCCH uplink and downlink CCE sharing function". Afterwards, the edge computing node collects the CCE utilization index of each cell, determines the current CCE load of the corresponding cell based on the CCE utilization index, and generates a channel occupancy adjustment instruction when the current CCE load of the cell is greater than the set threshold; the channel occupancy adjustment instruction is sent to the base station, so that the base station adjusts the channel parameters based on the channel occupancy adjustment instruction. This step takes into account that PDCCH is an important influencing factor for load regulation.

[0123] That is to say, the embodiment of the present invention introduces a CCE utilization evaluation mechanism. On the basis of the service channel load evaluation, the CCE utilization factor is taken into consideration. When the CCE utilization rises, the probability of signaling congestion is reduced by increasing PDCCH resources, thereby improving the stability of the load dynamic control mechanism and improving user perception.

[0124] See also Figure 2 , is a flowchart of a training method for a machine learning model provided by an embodiment of the present invention, the method is applied to a centralized computing node, and includes:

[0125] Step 201: Initialize model parameters of the machine learning model;

[0126] Step 202: Obtain a training data set and a test data set;

[0127] Step 203: Based on the training data set, the machine learning model after initializing the model parameters is trained for multiple rounds, wherein each round of training includes: initializing the measurement variables of the training data set, reading the training data set and its labels, and storing the label values ​​of the labels; traversing the training data set, during the traversal process, calculating the characteristic values ​​using the model parameters, and storing the calculated characteristic values ​​as the characteristic values ​​output by this batch of training data, using the loss function to calculate the loss value, calculating the gradient based on the loss value, performing back propagation, and updating the machine learning model function, updating the measurement variables, and storing the loss value as the calculation variable; reading the stored label values ​​and the characteristic values ​​output by this batch of training data, calculating the accurate number of this batch of training data based on the label values ​​and the characteristic values, and updating the accurate number and the total number into the measurement variables, and superimposing the original values ​​to obtain the calculation variables including the accurate number, the total number and the loss value; repeating the above process until all the theoretical training is completed, and the accuracy of the corresponding training results is output;

[0128] Step 204: When the accuracy reaches a threshold, determine that the machine learning model is a trained model.

[0129] Optionally, obtaining a training data set and a test data set includes: obtaining traffic data of each cell in a high-load protection area; merging the obtained traffic data by cell to obtain a cell-level merge table; performing operations on the fields of the cell-level merge table to generate new fields as a basis for labeling and training; normalizing rate-related fields according to the network configuration of each cell; and grouping the normalized data of each cell at different time periods to obtain a training data set and a test data set.

[0130] The step of acquiring the training data set and the test data set further includes: reorganizing and labeling the training data set and the test data set using a sliding window, specifically including:

[0131] A training data set and a test data set are grouped using sliding windows respectively; tensors and scalars corresponding to the grouped training data set and test data set are determined; the tensors and scalars are grouped according to a corresponding set batch size, wherein the training data set is shuffled before grouping, and the test data set is not shuffled before sub-grouping.

[0132] In the embodiment of the present invention, a federated learning technology architecture combining centralized computing nodes and edge computing nodes is adopted. The centralized computing nodes are mainly responsible for processing large amounts of data, training machine learning models and outputting parameters, while the edge computing nodes are mainly responsible for quickly extracting data features during the guarantee period and quickly generating base station adjustment instructions. The advantages of both are fully combined so that the results of machine learning can be applied to scenarios such as communication guarantee with high real-time requirements.

[0133] See also Figure 3 , is a flow chart of an application example of a method for dynamically adjusting a wireless network load provided by an embodiment of the present invention. The embodiment of the present invention provides processes such as labeling, model training, model parameter distribution, script generation and implementation during the implementation process. The method includes:

[0134] Step 301: The centralized computing node initializes the model parameters of the machine learning model; specifically, historical statistical data can be collected to train the model parameters, and the training includes the process of training data sets and test data and their annotation, model parameter training, etc.

[0135] This step realizes the close integration of the learning process of the machine learning model and various parameter configurations with the dynamic adjustment of the wireless network load. The specific implementation process is detailed in the subsequent embodiments.

[0136] Step 302: The centralized computing node deploys the model parameters of the trained machine learning model to the edge computing nodes in the large traffic guarantee area.

[0137] This step defines the message exchange mechanism between the centralized computing node and the edge computing node. Based on the training in step 301, the centralized computing node archives and manages the model parameters. The model is divided according to the BBU room, frequency band and bandwidth where the edge computing node is located, that is, the cell targeted by each model must belong to the same BBU room, the same frequency band and the same bandwidth. For models containing multiple cells, through multiple parameter optimizations and before guarantee, the centralized computing node sends the model parameters corresponding to the guarantee cell to the edge computing node.

[0138] Step 303: The edge computing node uses a machine learning model to analyze the traffic load data in real time, outputs traffic decision point features, and converts the traffic decision point features into base station parameter adjustment commands; and sends the base station parameter adjustment commands to the base station; the base station executes the script of the base station parameter adjustment commands.

[0139] This step defines the mapping relationship between the feature and the base station parameter adjustment command. That is to say, the edge computing node reads the pth file in the received model parameters for the model corresponding to all the cells that need to be protected, loads the model parameters into the memory, and obtains the machine learning model of the corresponding cell. And use the machine learning model to process the traffic load data of the corresponding cell in real time to obtain the characteristic value of the traffic decision point; convert the characteristic value of the traffic decision point into a base station parameter adjustment command; send the base station parameter adjustment command to the base station so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command; evaluate the real-time data obtained after adjusting the base station parameters to determine the threshold adjustment instruction; send the threshold adjustment instruction to the centralized computing node so that the centralized computing node performs the corresponding threshold adjustment based on the threshold adjustment instruction, and retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node.

[0140] Step 304: The edge computing node decides whether to adjust the CCE parameters according to the current CCE load and generates an adjustment script; the edge computing node sends the base station parameter adjustment to the base station; the base station executes the script. This step considers that PDCCH is an important factor affecting load regulation.

[0141] In this step, before the guarantee starts, as the default value, the "PDCCH channel occupied symbols" of the cell is set to 2 symbols, the "PDCCH channel occupied RB number" is set to 0 (i.e. occupying the full bandwidth), and the "PDCCH uplink and downlink CCE sharing function" is turned on. First, the edge computing node collects the CCE utilization index of each cell, and determines the current CCE load of the corresponding cell based on the CCE utilization index; when the current CCE load of the cell is greater than the set threshold, a channel occupancy adjustment instruction is generated; the channel occupancy adjustment instruction is sent to the base station, so that the base station adjusts the channel parameters based on the channel occupancy adjustment instruction.

[0142] It should be noted that step 304 is an optional step.

[0143] Step 305: The edge computing node collects data in real time and evaluates the implementation effect; outputs an adjustment request for the relevant threshold in the labeling data algorithm based on the evaluation result, notifies the centralized computing node to adjust the threshold, and retrains the model on this basis, and sends the model parameters to the edge computing node.

[0144] This step defines the dynamic adjustment mechanism of the threshold in the marking algorithm.

[0145] In this step, the edge computing node can collect data in real time and evaluate the implementation effect based on step 303 or step 304.

[0146] In the embodiments of the present invention, based on the relevant technology, the trend of network load changes in the future period is predicted by using machine learning algorithms based on historical traffic data, which is often used for wireless network planning. The means taken are time-consuming actions such as adding cells or expanding capacity, which fail to meet the scenarios with high real-time requirements such as emergency support. In the embodiments of the present invention, the results of machine learning can be applied to dynamically adjust the network in such scenarios to meet the technical problem of real-time adjustment of 5G base station capacity, and the label output by the machine learning algorithm is mapped to the dynamic adjustment instruction, thereby dynamically adjusting the soft capacity of the 5G base station in real time and improving the technical effect of customer perception.

[0147] In an embodiment of the present invention, a method for dynamically adjusting wireless network load combined with machine learning is proposed, which adopts a federated learning technology architecture that combines centralized computing nodes and edge computing nodes. The centralized computing nodes are mainly responsible for processing large amounts of data, training machine learning models, and outputting parameters. The edge computing nodes are mainly responsible for quickly extracting data features during the guarantee period and quickly generating base station adjustment instructions. The advantages of both are fully combined, so that the results of machine learning can be applied to scenarios such as communication guarantee with high real-time requirements.

[0148] In the embodiments of the present invention, it is proposed to use the temporal distribution of multiple wireless indicators such as PRB utilization as a "photo" of two-dimensional pixel distribution according to the characteristics of 5G wireless communication operation, and abstract the corresponding load adjustment instructions to be taken as "labels", so as to understand wireless network indicators from a new perspective, which is conducive to the application of various algorithm models used for image recognition processing in modern machine learning.

[0149] The embodiment of the present invention proposes to apply load assessment logic to the algorithm for marking training set data and test set data, thereby achieving rapid data marking, reducing manual intervention, and improving marking accuracy.

[0150] The embodiment of the present invention proposes to set some threshold parameters in the marking algorithm to be adjustable in view of the continuous changing characteristics of the wireless network during the protection process, so that the marking algorithm itself can dynamically change according to the changes in the wireless network, thereby promoting the dynamic adjustment of load instructions to adapt to network changes and achieve the purpose of more effective protection of user perception.

[0151] The embodiment of the present invention proposes that in terms of network load adjustment instructions, in addition to basic channels, scheduling, MIMO configuration and other aspects, factors of the control channel - PDCCH are also considered, and basic network access capabilities are guaranteed by adjusting PDCCH parameters.

[0152] See also Figure 4 , which is a schematic diagram of a computing architecture established in combination with a 5G wireless network architecture provided in an embodiment of the present invention.

[0153] As shown in the figure, first of all, the network software and architecture used in this embodiment take the 5G wireless network as an example. The network element equipment involved includes: 5G wireless network management and BBU (a computer room with multiple BBUs concentrated), and multiple AAU / RRUs under each BBU. Based on the wireless network architecture, centralized computing nodes and edge computing nodes are deployed (one is taken as an example, and it is not limited to this in actual applications). In addition, in order to handle exceptions that occur during instruction execution, this embodiment also involves centralized monitoring nodes in the figure. The entire network and software architecture is shown in the figure. Figure 4 As shown.

[0154] Centralized computing nodes can usually be deployed at the provincial or municipal level, and do not require high network latency. They are responsible for the collection and storage of training and test data sets, the processing and conversion of training and test data sets, the maintenance of MLP model parameters, the training of model parameters (which may include temporary model training after receiving edge computing node instructions during the guarantee period), and the output of model parameters to edge computing nodes.

[0155] The edge computing node is usually deployed in the computer room where the BBU is located, connected to the BBU through the STN-A device, and is in the same local area network as the BBU's main control board, and is directly connected to the proximal maintenance network management of the main control board. Based on the above structure, the edge computing node sends instructions to the BBU through the proximal maintenance network management, and the network delay with the BBU reaches the millisecond level. The edge computing node is responsible for using the machine learning model to analyze the traffic load data in real time, output the traffic decision point label; convert the traffic decision point label into a base station parameter adjustment script; it can also improve the base station parameter adjustment script in combination with the CCE load situation; and send the improved base station parameter adjustment script to the BBU; and during the guarantee period, send the machine learning model retraining instruction to the centralized computing node. In this embodiment, the BBU is responsible for the execution of the script.

[0156] The centralized monitoring node is responsible for handling exceptions during the execution process. Under normal circumstances, the probability of participation is very low.

[0157] In this embodiment, multiple databases (tables) are also deployed on the centralized computing node and the edge computing node to assist in the process of computing and instruction execution. For example, the centralized computing node deploys a parameter model database, a model number and cell information index database, a threshold parameter database, etc., and the edge computing node deploys a parameter model database, an instruction execution (success) record table database, a model number and cell information index database, a database storing an instruction execution failure alarm table, a threshold parameter database, etc. However, in specific applications, it is not limited to this.

[0158] Secondly, the machine learning model

[0159] The machine learning in this embodiment adopts an MLP (multi-layer perception) model, and the software architecture of the model is shown in Figure 5 (1) and Figure 5 (2). Figure 5 (1) is a schematic diagram of the hierarchical structure of an MLP model provided in an embodiment of the present invention; Figure 5 (2) is a schematic diagram of the processing of an MLP model provided in an embodiment of the present invention.

[0160] As shown in Figure 5(1), the MLP model consists of the following steps: input data (7×120), data flattening (840), input layer (840×64), input layer activation (ReLU), input layer dropout (probability 0.2), hidden layer (64×64), hidden layer activation function (ReLU), hidden layer dropout (probability 0.3), output layer (4×64), output data (16 label values), etc.

[0161] As shown in FIG5(2), in this embodiment, the MLP model mainly includes an input layer, a hidden layer, and an output layer, which correspond to three matrix operations (weights), with sizes of 840×64, 64×64, and 64×28, respectively, and a bias is added on this basis; after these three layers of processing, an array (tensor) of 28 elements is output, and the index corresponding to the maximum value of these 28 elements is taken as the output label.

[0162] The model can be described in Python language PyTorch library:

[0163] Sequential(

[0164] (0):Flatten(start_dim=1,end_dim=-1)

[0165] (1):Linear(in_features=840, out_features=64, bias=True)

[0166] (2): ReLU()

[0167] (3):Dropout(p=0.3,inplace=False)

[0168] (4):Linear(in_features=64, out_features=64, bias=True)

[0169] (5):ReLU()

[0170] (6):Dropout(p=0.2,inplace=False)

[0171] (7):Linear(in_features=64, out_features=28, bias=True) )

[0173] Finally, storage of machine learning models

[0174] As mentioned earlier, a machine learning model consists of multiple layers, each of which contains weights and biases, as well as some other parameters. These parameters are stored in pth files (binary format).

[0175] The centralized computing node and the edge computing node both store the pth file of the relevant model parameters and index them with the parameter model database, as shown in Table 1, which is the parameter model database. In addition, the centralized computing node and the edge computing node also store the index table of the model number and the cell information, as shown in Table 2, which is the index table of the model number and the cell information. It should be noted that Tables 1 and 2 are only for example, and are not limited to this in practical applications.

[0176] Table 1

[0177]

[0178]

[0179] Table 2

[0180] Model Number Base station number Cell number 1 7912345 0 1 7912346 1 2 7912345 1 2 7912347 0 3 7912347 1 3 7912347 2 …… …… ……

[0181] Based on the above, the present invention also provides two typical embodiments, taking model training and dynamic adjustment for several cells in the 3.5G frequency band and the 2.1G frequency band as examples. Most of the steps in the two embodiments are consistent, and a few steps are different, which will be specifically described in the following description.

[0182] First, the centralized computing node initializes the model parameters of the machine learning model; specifically, historical statistical data can be collected to train the model parameters, and the training includes the labeling of the training data set and the test data, the training of the model parameters, etc. This step specifically includes steps 111) to 119).

[0183] 111) First, in the centralized computing node and the edge computing node, an MLP machine learning model is established according to the structural diagrams of Figure 5(1) and Figure 5(2). The global parameters (hyperparameters) of the model are shown in Table 3. Table 3 is the hyperparameters of the MLP machine learning model. It is only an example and is not limited to this in practical applications.

[0184] Table 3

[0185]

[0186]

[0187]

[0188] 112) The centralized computing node collects the traffic data of each cell in the high-load protection area.

[0189] Among them, the source of data collection: the model training data before the guarantee is collected from the centralized network management, and the temporary training data during the guarantee period comes from the report of the edge computing node. The original data is a file stored in csv format. The embodiment of the present technical solution uses the pandas library of python to read the csv file and stores it in the memory in DataFrame format for subsequent processing. The collected data is divided into three tables (corresponding to three csv files): total throughput, number of users, and wireless resource usage. The data examples are shown in Table 4, Table 5, and Table 6 respectively. The granularity of the data table record is 1 second, and all records are made for a single cell, with a total of 9 indicators recorded. In this embodiment, considering the existing network configuration, a 5G cell is only configured with one NR DU cell and a corresponding TRP. Therefore, Table 4, Table 5 and Table 6 can be merged according to the cell dimension, and the three DataFrames can be merged into one using a python program. Among them, Table 4 shows an example of a total throughput monitoring table, Table 5 shows an example of a user number monitoring table; Table 4 shows an example of a wireless resource usage monitoring table.

[0190] Table 4

[0191]

[0192]

[0193] Table 5

[0194] Collection time Total number of users in the community 2024-08-05 20:59:44(263) 86 2024-08-05 20:59:45(264) 85 2024-08-05 20:59:46(264) 87 …… ……

[0195] Table 6

[0196]

[0197] 113) The centralized computing node merges the collected data in units of cells. The basis for the merger is the "collection time" field in Table 4, Table 5 and Table 6. The alignment method is as follows:

[0198] Assume that the data tables (DataFrame) recorded in Table 4, Table 5, and Table 6 are T3, T4, and T5 respectively, and the "collection time" field of the first data in the three tables satisfies T3[0]['time']≤T4[0]['time']≤T5[0][time'], then the calculated indexes i3 and i4 satisfy:

[0199]

[0200] This completes the time alignment by taking numbers from index i3, i4, and 0 respectively for T3, T4, and T5. In this way, a new table T is formed for each cell in the protection area. For high load evaluation, the MAC layer rate is concerned but not the RLC layer rate. Therefore, the final format of T is shown in Table 7, which shows the field format of the cell-level merge table.

[0201] Table 7

[0202]

[0203] 114) Operate the fields of the cell-level merge table T to generate new fields as the basis for labeling and training. The operation rules are as follows:

[0204] For the i-th row of data in T, we have

[0205] prb_downlink_capacity←slots_per_second1·downlink_percentage·Number_PRBs

[0206] prb_uqlink_capacity←slots_per_second·uplink_percentage·Number_PRBs

[0207] T[i][′Downlink PRB utilization rate′]←T[i][′Total number of downlink RBs used′] / prb_downlink_capacity·100

[0208] T[i][′Uplink PRB utilization′]←i[i][′Total number of uplink RBs used′] / prb_uplink_capacity·100

[0209]

[0210] The parameters related to the calculation of PRB utilization, including the time slot period (number of time slots per second), the downlink time slot ratio, the uplink time slot ratio, and the number of PRBs scheduled per time slot, are defined by the 3GPP 38.101 series of protocols. Typical configuration values ​​of the existing network are shown in Table 8. Table 8 PRB utilization calculation related parameters. In this embodiment, two configurations corresponding to two embodiments are taken as examples.

[0211] Table 8

[0212]

[0213] 115) The four rate-related fields are normalized according to the network configuration of the cell, and the normalization operation rules are as follows:

[0214]

[0215]

[0216]

[0217] The normalization coefficients for typical network configurations in the above calculation formula are shown in Table 9, which represents the normalization coefficients for typical network configurations.

[0218] Table 9

[0219]

[0220] 116) Group the data of different time periods of each cell after the previous step to form training data and test data of different models.

[0221] First, the models are divided. For the different network configurations described in Tables 8 and 9, at least one MLP model should be defined for each network configuration. If the station heights and surrounding wireless environments of the cells in each network configuration vary greatly, they should be further subdivided into different models for separate training and evaluation.

[0222] Then, the training data set and the test data set are divided. The general principle of the division is to maintain the daily integrity of the data, that is, the data of a community on the same day is either placed in the training data set or in the test data set. The specific operation process is to read the community data file corresponding to a model in the protection area, and Figure 6 Write the flowchart into the training data set or test data set, Figure 6 A flowchart of writing a single indicator file into a training data set or a test data set is provided in an embodiment of the present invention.

[0223] 117) The training data set and the test data set are reorganized and labeled using a sliding window. The process is as follows: Figure 7 FIG. 1 is a schematic diagram of a data reorganization and marking process provided by an embodiment of the present invention, which specifically includes:

[0224] First, the data set is grouped using a sliding window. In two embodiments of the present invention, the sliding window size is 120 (i.e., 2 minutes) and the sliding step size is 10 (i.e., 10 seconds). Figure 7As shown in the figure, the input is a data table with a dimension of n×7, corresponding to the training set or test set data. The data table is divided into small matrices of 10×7. The sliding window starts reading from the head of the input data table, reading 12 small matrices each time, outputting a 120×7 matrix, and upgrading the matrix to obtain a 1×120×7 tensor; after reading, the sliding window slides forward 10 (that is, the next time it starts reading from the next small matrix). This process continues until the sliding window covers the last complete small matrix (with 10 rows of data).

[0225] Then, each 1×120×7 tensor outputted above is labeled. The initial labeling algorithm is described as follows:

[0226]

[0227]

[0228] The threshold initial values ​​described in the above marking algorithm (consistent in the two embodiments) are set as shown in Table 10 in the two embodiments. In the embodiment of the technical solution of the present invention, only three thresholds related to PRB utilization are adjusted, namely and T uplink，prb Therefore, on this basis, the minimum and maximum values ​​of these three thresholds are provided, but the minimum and maximum values ​​of the other six thresholds are not provided. During the adjustment process, the following formula must also be ensured:

[0229]

[0230] As shown in Table 10, Table 10 shows the threshold initial value, minimum value and maximum value of the marking algorithm in the two embodiments.

[0231] Table 10

[0232]

[0233]

[0234] Table 10 lists the initial value, minimum value and maximum value of the three thresholds related to the PRB utilization rate among the nine thresholds. These thresholds can be dynamically adjusted according to the changes in base station parameters and load indicators, and the models can be re-labeled and trained. In the centralized computing stage and edge computing nodes, these three threshold values ​​of each model are stored, and their formats are shown in Table 11. Table 11 shows the threshold parameter format stored in the centralized computing node and the edge computing node.

[0235] Table 11

[0236]

[0237] 118) The tensors {X1, X2, ...} and scalars {Y1, Y2, ...} corresponding to the training data set and the test data set output in the previous step are grouped according to the "batch size batch_size" setting value (16) shown in Table 3. The training data set is shuffled before grouping, while the test data set is not shuffled. The two embodiments of this embodiment are implemented by calling the DataLoader class in PyTorch, and the calling statements of the training data set and the test data set are as follows:

[0238] dataloader_train=DataLoader(dataset_train,batch_size=batch_size,shuffle=True)

[0239] dataloader_test=DataLoader(dataset_test,batch_size=batch_size,shuffle=False)

[0240] In the two embodiments, the tensor sizes and the number of tensors before and after grouping are respectively shown in Table 12, which shows the tensor sizes and the number of tensors before and after grouping.

[0241] Table 12

[0242] Example Example 1 Example 2 Frequency band 3.5G 2.1G bandwidth 100MHz 20MHz Before grouping - training dataset - data - tensor size 30487×1×120×7 25791×1×120×7 Before grouping - training data set - label - tensor size 30487×1 25791×1 Before grouping - training data set - data - number of tensors 1 1 Before grouping - training data set - label - number of tensors 1 1 Before grouping - test dataset - data - tensor size 25824×1×120×7 17229×1×120×7 Before grouping - test data set - label - tensor size 25824×1 17229×1 Before grouping - test data set - data - number of tensors 1 1 Before grouping - test data set - label - number of tensors 1 1 After grouping - training data set - data - tensor size 16×1×120×7 16×1×120×7 After grouping - training data set - label - tensor size 1906 1612 After grouping - training data set - data - number of tensors 16×1 16×1 After grouping - training data set - label - number of tensors 1906 1612 After grouping - test dataset - data - tensor size 16×1×120×7 16×1×120×7 After grouping - test data set - label - tensor size 1614 1077 After grouping - test data set - data - number of tensors 16×1 16×1 After grouping - test data set - label - number of tensors 1614 1077

[0243] 119) The centralized computing node uses the training data set and test data set output in the previous step to train the model parameters and evaluate the training effect. The corresponding process is as follows Figure 8 , Figure 9-1 , Figure 9-2 , Fig.10 and Fig.11 shown.

[0244] Figure 8 The overall algorithm flow chart of an MLP model parameter training provided by an embodiment of the present invention is as follows: Figure 8 As shown:

[0245] First, the node initializes the model parameters, further reads the training data set and the test data set, and performs multiple rounds of training and evaluation on the training data set. That is, it is necessary to determine whether all epochs are completed. If so, the final result is output. Otherwise, a training is performed on the training data set. The process corresponds to Figure 9-1 ; After that, the model is used to calculate the accuracy of the test data set, and the process corresponds to Fig.10 .

[0246] In each round, firstly, a training is performed on the training data set, corresponding to Figure 9-1 After that, the accuracy of the test data set is calculated using the model. The process corresponds to Fig.10 During this training process, a "measurement variable" data storage is established to store results such as accurate numbers, total numbers, and loss values.

[0247] Figure 9-1 for Figure 8 Detailed expansion of "Perform a training run on the training dataset" in the overall flowchart.

[0248] First, initialize the measurement variables of the training data set, that is, set the results of each parameter to 0. Read the training data set and its labels, and temporarily store the labels in the "label value" data storage.

[0249] Secondly, the training data set is traversed. During the traversal process,

[0250] First, use the model parameters of the model to calculate the eigenvalues, and output the results to the "Eigenvalues ​​of this batch of training data output" data storage. Use the loss function to calculate the loss value. Secondly, calculate the gradient based on the loss value, perform backpropagation, and update the model function, update the measurement variable - the loss value, read the "label value" and "Eigenvalues ​​of this batch of training data output" data storage contents, use the values ​​of these two data storages to calculate the exact number of this batch of training data sets, and update the exact number and total number to the measurement variable, and superimpose the original value. The specific process is as follows: Figure 9-1 The steps are shown in the figure and will not be repeated here.

[0251] in, Figure 9-2 for Figure 9-1 The expanded flowchart of "Calculate the exact number of training data sets in this batch" in the flowchart.

[0252] First, read the contents of the data storage of "label value" and "feature value output of this batch of training data", compare the label value and the feature value output of the training data. Calculate the number of consistent values ​​between the two, update the measurement variables, including the exact number and the total number, and superimpose the original values.

[0253] Fig.10 for Figure 8 A detailed flowchart of "Using the model to calculate the accuracy of test data" in the overall flowchart.

[0254] First, initialize the measurement variables of the test data set, including the exact number and the total number, and set them to 0. Read the test data set and its corresponding labels, and temporarily store the labels in the "label value" data storage. Traverse the test data set.

[0255] During the traversal process, first, the eigenvalues ​​are calculated using the current parameters of the model, and the results are output to the "Eigenvalues ​​of this batch of training data output" data store. The contents of the "label value" and "Eigenvalues ​​of this batch of training data output" data stores are read, and the values ​​of these two data stores are used to calculate the exact number of this batch of test data sets. The process is detailed in Fig.11 , and update the accurate number and total number to the measurement variable, and superimpose the original value. The specific implementation process is detailed in Fig.10 .

[0256] Fig.11 for Fig.10 The expanded flowchart of "Calculate the exact number of test data sets in this batch" in the flowchart.

[0257] First, read the contents of the data storage for "label value" and "feature value output of this batch of training data", and compare the label value and the feature value output of the training data. Calculate the number of consistent values ​​between the two. Update the measurement variables, including the exact number and the total number, and superimpose the original values.

[0258] It should be noted that the model training parameters of the two embodiments, such as the number of training rounds, learning rate, etc., are shown in Table 1.

[0259] The initial training loss rate and accuracy change curves of the two embodiments are as follows: Fig.12 and Fig.13 shown.

[0260] like Fig.12 and Fig.13 In the example, train loss refers to the training loss, train acc refers to the training accuracy, and testacc refers to the test accuracy. An epoch refers to the number of times all data are trained once, and the number represented by epoch refers to the total number of rounds that all data are trained.

[0261] The final training loss rate and accuracy of the initial training output of the two embodiments are shown in Table 13. Table 13 shows the final training loss rate and accuracy of the initial training output of the two embodiments.

[0262] Table 13

[0263] Example Example 1 Example 2 Frequency band 3.5G 2.1G bandwidth 100MHz 20MHz Training dataset loss rate 0.21484241846122962 0.1589065732180247 Training data set accuracy 0.9187522550595336 0.9385832267069908 Test data set accuracy 0.8639637546468402 0.8985431539845609

[0264] Optionally, in this embodiment, the centralized computing node deploys the model parameters of the trained machine learning model to the edge computing node in the large traffic guarantee area. Specifically including 211) to 213):

[0265] 211) The centralized computing node completes the archiving and indexing update of the model parameters.

[0266] First, the centralized computing node archives and manages the model parameters based on the training in 101). The model is divided according to the BBU room, frequency band and bandwidth where the edge computing node is located, that is, the cell targeted by each model must belong to the same BBU room, the same frequency band and the same bandwidth. In an embodiment of the technical solution of the present invention, the centralized computing node uses a function similar to "torch.save(model.state_dict(),' / XXX / model_weights.pth')" in PyTorch to complete the archiving of model parameters.

[0267] After archiving, update the contents of the two index tables, Table 1 and Table 2.

[0268] 212) For a model containing multiple cells, if the accuracy of the training data set or the accuracy of the test data set is lower than 0.8 after multiple parameter optimizations, the centralized computing node adds more statistical data for the cells in the model whose cell indicator statistical file has a statistical length of less than 3 days, and then follows the aforementioned steps 116) and Figure 6 The algorithm divides the training data set and the test data set, and then retrains the model and adjusts the hyperparameters according to the aforementioned steps 117) to 119). If after multiple adjustments, the "accuracy of the training data set or the test data set is not less than 0.8" cannot be met, the centralized computing node splits the model according to the cell, and reprocesses the data and trains the model according to step 101) until the above conditions are met or only one cell is contained in a single model. After adjustment, the model and cell information index table described in Table 2 need to be updated.

[0269] 213) Before key protection, the centralized computing node sends the parameters of the corresponding model of the protection cell to the edge computing node.

[0270] First, the centralized computing node queries Table 1 and Table 2 to obtain the parameter storage location of the model corresponding to each cell.

[0271] Then, the centralized computing node loads the model parameters into the memory. In the embodiment of the present invention, a function similar to "model.load_state_dict(torch.load(' / XXX / model_weights.pth'))" in PyTorch is used to complete the process.

[0272] After that, the centralized computing node completes the model parameter transmission with the edge computing node. An example of a signaling interaction diagram defining the model parameter transmission process is shown in Fig.14 shown. Fig.14 In the above example, the parameters of two models are updated. The update process of each model parameter is as follows:

[0273] The centralized computing node sends a model parameter sending request to the edge computing node, including the model number and the corresponding cell (cell 1, cell 2, ...); the edge computing node queries the cell list under the BBU room where the node is located to confirm whether the cells in the received model parameter sending request are all in the cell list; if they are all in the list, the edge computing node replies to the centralized computing node with a confirmation message including the model number; the centralized computing node queries the locally stored Table 1 to obtain the pth file corresponding to the model; the centralized computing node sends the pth file to the edge computing node; the edge computing node saves the pth file and updates the locally stored Table 1 and Table 2.

[0274] Optionally, in this embodiment, the edge computing node uses a machine learning model to analyze the traffic load data in real time, outputs traffic decision point features, and converts the traffic decision point features into base station parameter adjustment commands; and sends the base station parameter adjustment commands to the base station; the base station executes the script of the base station parameter adjustment commands. Specifically, it includes 311) to 318) as shown:

[0275] 311) First, the edge computing node reads the locally stored pth file for the models corresponding to all cells that need to be protected, and loads the model parameters into the memory.

[0276] 312) After the protection starts, the edge computing node collects data in the formats of Table 4, Table 5, and Table 6 in real time for each cell that needs to be protected, merges the data in the three tables every 2 minutes (120 seconds), and calculates the 7 fields of the 120 data points with reference to the algorithms of steps S114 and S115 to form a 1×120×7 tensor.

[0277] 313) The edge computing node uses the model parameters corresponding to each protected cell to calculate the 1×120×7 tensor output in the previous step and output its eigenvalue Y i .

[0278] 314) Edge computing node uses 117) to calculate Y t ←C upllnk +4·C downllnk The reverse method is to infer C uplink and C downlink The specific calculation formula is:

[0279] C uplink =Y i mod4

[0280]

[0281] Because Y t The value range of C is 0 to 27.uplink The value range is 0 to 3, C downlink The value range is 0 to 6.

[0282] 315) Set network parameters according to the rules of Table 14 to Table 17 below, where Table 14 represents C uplink Table 15 shows the corresponding network parameter setting rules. uplink Corresponding network parameter setting rules (continued); Table 16 shows C downllnk Corresponding to the network parameter setting rules, Table 17 shows C downlink Corresponding network parameter setting rules (continued).

[0283] Table 14

[0284]

[0285] Table 15

[0286]

[0287]

[0288] Table 16

[0289]

[0290] Table 17

[0291]

[0292]

[0293] 316) The edge computing node compares the current C uplink and C downlink The value of C output by 315) uplink and C downlink The comparison process includes:

[0294] First, the edge computing node maintains a command execution (success) record table, the format of which is shown in Table 18. The second and third columns in the table respectively represent the C based on which the command was issued at that time. uplink and C downlink The fourth column is in json format, recording the values ​​of various parameters updated according to the execution results returned by the base station, where <key1> 、 <key2>, ...corresponding to the parameter names defined in Tables 14 to 17, (in the i-th row) <valuei1> 、 <valuei2>, ... corresponding to this moment <key1> 、 <key2>, ... parameter values. Among them, Table 18 shows the instruction execution (success) record table

[0295] Table 18

[0296]

[0297]

[0298] Then, at a certain moment, assuming that the (last row) most recent C recorded for a certain cell in Table 18 is uplink and C downlink The values ​​are and The values ​​output by 315 are and The parameter value json recorded in the last row of Table 18 is P i . Assume that the target generation modification parameter value json is Q i , an algorithm description of its generation process is as follows:

[0299]

[0300] 317) The edge computing node outputs the Q t Converted into parameter modification instructions that can be understood by the (base station) BBU. The parameters listed in Tables 14 to 17 are divided into three types, including numerical type, enumeration type, primary switch type, and secondary switch type. Table 19 gives an example of the mapping relationship between some parameters of a certain manufacturer and the parameters of the network management modification instruction; Among them, Table 19 shows an example of the mapping relationship between some parameters of a certain manufacturer and the parameters of the network management modification instruction.

[0301] Table 19

[0302]

[0303]

[0304]

[0305] The conversion process in this embodiment specifically includes:

[0306] First, the edge computing node targets Q t Each key-value pair

[0307] Generate a corresponding single-cell, single-parameter modification script, and temporarily store the script in the memory. For the manufacturer network management parameters used in the embodiment, the BBU name is BBU_NAME (which has a one-to-one correspondence with the base station number described in Table 2, as the unique identifier of the base station), and the NR DU cell number is NRDUCELLID (sometimes NRCELLID, which uniquely identifies a cell), the mapping rules for modifying the script are as follows:

[0308] For numeric and enumeration type parameters, query Table 19 to obtain The corresponding command name is Instruction, the first-level IE is Key, and the single-parameter modification script is

[0309] For the parameters of the primary switch type, query Table 19 to obtain The corresponding command name is Instruction, and the first-level IE (switch name) is Key. The corresponding switch value is SWITCH (value is ON or OFF), then the single parameter modification script is "MOD <instruction>:NRDUCELLID=’XXX’, <key> ＝ <switch>; {<BBU_NAME>}";

[0310] For the parameters of the secondary switch type, query Table 19 to obtain The corresponding command name is Instruction, the first-level IE is Key, and the second-level IE (switch name) is SWITCH_NAME. The corresponding switch value is SWITCH_VALUE (value is 0 or 1), then the single parameter modification script is "MOD <instruction>:NRDUCELLID=’XXX’, <key>=<SWITCH_NAME> -<SWITCH_VALE> ; {<BBU_NAME>}".

[0311] Then, based on this, the edge computing node merges the single parameter modification script according to the structure of the base station network management command, outputs the single cell and multi-parameter modification script, and temporarily stores it in the memory. The merging rules are:

[0312] The two merged single parameter modification scripts must belong to the same base station and the same BBU, and use the same command name;

[0313] If both scripts are of the secondary switch type and the primary IE is the same, they will be merged at the IE level, such as "MOD <instruction>:NRDUCELLID=’XXX’, <key>=<SWITCH_NAME1> -<SWITCH_VALE1> ; {<BBU_NAME>}" and "MOD <instruction>:NRDUCELLID=’XXX’, <key>=<SWITCH_NAME2> -<SWITCH_VALE2> ; {<BBU_NAME>}" two commands are combined into "MOD <instruction>:NRDUCELLID=’XXX’, <key>=<SWITCH_NAME1> -<SWITCH_VALE1> &<SWITCH_NAME2> -<SWITC H_VALE2> ; {<BBU_NAME>}";

[0314] In other cases, merging is done at the command level, such as {<BBU_NAME>}" and "MOD <instruction>:NRDUCELLID=’XXX’, <key2> ＝ <switch>; {<BBU_NAME>}" The two commands are combined into

[0315]

[0316] The merging process is to read single-parameter scripts sequentially and merge them step by step. In the middle, there is a situation where multi-parameter scripts and single-parameter scripts are merged. The principles are the same as the above two.

[0317] 318) The edge computing node sends the base station parameter modification instruction to the corresponding BBU, and the BBU returns the result after executing the instruction, and the edge computing node updates table 18. An example of the instruction execution signaling interaction between the edge computing node and the BBU (BBU1 and BBU2 are taken as examples in this embodiment) defined in this step is as follows Fig.15 shown.

[0318] The signaling interaction process is described as follows:

[0319] First, the edge computing node sends a multi-parameter modification instruction and a (unmerged) single parameter modification instruction to the relevant BBU (for example, BBU1, or BBU2, or to BBU1 and BBU2 at the same time). The instruction contains a cell list and a modification script. If multiple cells are involved, the modification scripts of multiple cells are merged and sent together.

[0320] Then, after receiving the instruction, the corresponding BBU executes the modification script.

[0321] Then, the BBU returns the command execution result. The execution result is fed back for each cell, indicating whether the execution is successful. If the result is a failure, the detailed reason for the execution failure is also returned.

[0322] Then, the edge computing node decides whether to re-execute the failed instruction based on the instruction type. If it is a single parameter modification instruction, it will not be re-executed regardless of success or failure; if it is a multi-parameter modification instruction that fails to execute in some or all cells, the multi-parameter modification instruction will be split into multiple single parameter modification instructions (actually by searching the memory for the single parameter modification instructions before merging), and then these single parameter modification instructions will be sent to the BBU separately (the signaling contains the cell information of the failed execution), and the BBU will execute these instructions and return the execution results.

[0323] Then, the edge computing node integrates the results of all parameter modification instruction executions and updates the instruction execution (success) record table of each cell in Table 18, that is, adds one record to Table 18 to record the successful execution time of the instruction and the cell information. The last column "parameter value" is updated in combination with the previous record of this cell and the parameter values ​​in all successfully executed instructions.

[0324] Finally, all the single parameter modification instruction records that failed to execute are written into the instruction execution failure alarm record table, and its format is shown in Table 20. The last column "Whether to restore" is set to "No" when it is first added, and it is changed to "Yes" only after the manual processing is completed. Among them, Table 20 represents the instruction execution failure alarm record table.

[0325] Table 20

[0326]

[0327] Optionally, in this embodiment, the edge computing node decides whether to adjust the CCE parameters according to the current CCE load and generates an adjustment script; the edge computing node sends the base station parameter adjustment to the base station; the base station executes the script. Specifically including steps: 411) to 415):

[0328] 411) Before the guarantee starts, as the default value, set the cell "PDCCH channel occupied symbols" to 2 symbols, set the "PDCCH channel occupied RB number" to 0 (i.e. occupying the full bandwidth), and enable the "PDCCH uplink and downlink CCE sharing function".

[0329] 412) First, the edge computing node collects the CCE utilization index of each cell and stores the index as a DataFrame similar to step S101. The collection period is also 1 second, and the data format is shown in Table 21, which shows the CCE utilization index of the cell.

[0330] Table 21

[0331]

[0332] 413) The edge computing node calculates the total CCE utilization (%) at each moment. The calculation method in DataFrame is

[0333] T['Total CCE utilization (%)']

[0334] =T['Uplink DCI occupancy ratio (%)']+T['Downlink DCI occupancy ratio (%)']

[0335] +T[′SRSDCI occupancy ratio (%)′]+T[′CSIRSDCI occupancy ratio (%)′]

[0336] 414) The edge computing node evaluates the average total CCE utilization (%) of the last 5 minutes (300 seconds). If the indicator exceeds 60%, the edge computing node sends a "PDCCH channel occupied symbol number" adjustment instruction to the BBU where the cell is located, and adjusts the parameter to 3. The adjustment script is similar to "MOD NRDUCELLPDCCH:NrDuCellId=XXX,OccupiedSymbolNum=3SYM;".

[0337] 415) BBU executes the adjustment command and returns the result. Since the execution of the command will temporarily interrupt the cell service, once the command is successfully executed for a single cell, the command will not be executed before the guarantee is completed, so the entire step 304 is also skipped.

[0338] After the guarantee is completed, the "PDCCH channel occupied symbols" of all cells is restored to 2.

[0339] Optionally, in this embodiment, the edge computing node collects data in real time and evaluates the implementation effect; outputs an adjustment request for the relevant threshold in the labeling data algorithm according to the evaluation result, notifies the centralized computing node to adjust the threshold, and retrains the model on this basis, and sends the model parameters to the edge computing node. Specifically including: 511) to 516):

[0340] 511) First, based on step 303, the edge computing node calculates the value of each cell every 6 minutes (360 seconds) according to the calculation formula of step 114). and

[0341] The average values ​​within these 6 minutes, assuming that the two average values ​​of cell c1 are

[0342] and And the 3 Cs calculated in these 6 minutes downlink and C uplink Numeric value (because it is calculated every 2 minutes).

[0343] 512), the edge computing node calculates C in each cell downlink and C uplink If C downlink or C uplink If the three values ​​appear are all different, the last value is taken as the mode. Suppose the two modes of cell c1 are and The corresponding number of occurrences is and

[0344] 513) The edge computing node merges the modes of each cell obtained in 512), for example, the two modes of ci

[0345] and Then according to 512) query the average rate corresponding to the mode and

[0346] Assuming that the cell set in the model to be evaluated is {c1, c2, ...}, the merging and mapping algorithm is as follows:

[0347]

[0348] 513) Edge computing node compares two arrays and The value of each element and its corresponding target value generate the corresponding and T uplink，prb Three threshold adjustment instructions. Three typical frequency band / bandwidth configurations and The target value corresponding to each element and

[0349] As shown in Tables 22 to 24. Among them, Table 22 shows and The target value corresponding to each element and (3.5G / 100MHz);

[0350] Table 23 shows and The target value corresponding to each element and (2.1G / 20MHz); Table 24 and The target value corresponding to each element and (2.1G / 40MHz).

[0351] Table 22

[0352]

[0353] Table 23

[0354]

[0355] Table 24

[0356]

[0357] in, and T uplink，prb The generation algorithm of the adjustment instructions of the three thresholds is described as follows (where the adjustment instructions are defined by the json variable {'threshold_name':XXX,'direction_and_quantity':XX}, and the definition in Table 10 is T uplink，prb The minimum and maximum values ​​of the three thresholds are and

[0358] if satisfy and

[0359] If at least two of the following three conditions are met, the instruction {'threshold_name': 'direction_and_quantity':+2}:

[0360] Rate value

[0361] Rate value

[0362] Rate value

[0363] if satisfy If at least two of the following three conditions are met, the instruction {'threshold_name': 'direction_and_quantity':+2}:

[0364] Rate value

[0365] Rate value

[0366] Rate value

[0367] If T uplink，prb satisfy If at least two of the following three conditions are met, the instruction {'threshold_name':T uplink，prb ,'direction_and_quantity':+2}:

[0368] Rate value

[0369] Rate value

[0370] Rate value

[0371] if satisfy and

[0372] If at least two of the following three conditions are met, the instruction {'threshold_name': 'direction_and_quantity':-2}:

[0373] Rate value

[0374] Rate value

[0375] Rate value

[0376] if satisfy If at least two of the following three conditions are met, the instruction {'threshold_name': 'direction_and_quantity':-2}:

[0377] Rate value

[0378] Rate value

[0379] Rate value

[0380] If T upllnk，prb satisfy If at least two of the following three conditions are met, the instruction {'threshold_name':T upllnk，prb ,'direction_and_quantity':-2}:

[0381] Rate value

[0382] Rate value

[0383] Rate value

[0384] 514) The edge computing node merges the threshold adjustment instructions outputted by 513), that is, if both the raising instruction and the lowering instruction appear for the same threshold parameter, both instructions are deleted and the threshold is not adjusted. 'direction_and_quantity':+2} and {'threshold_name':

[0385] 'direction_and_quantity':-2} If two instructions appear at the same time, neither instruction will be executed.

[0386] 515) If the instruction set outputted in 514) is not empty, the edge computing node and the centralized computing node complete the update of the marking threshold parameters and the addition of test data in the model training through signaling interaction (including file transfer). The signaling interaction diagram is as follows: Fig.16 shown.

[0387] First, the edge computing node initiates a threshold modification instruction to the centralized computing node, including the corresponding model, the threshold to be updated, and the threshold parameter modification value. Then, the centralized computing node modifies the threshold parameter storage (storage format as Table 11) deployed on this node to modify the threshold parameter storage corresponding to the model. Then, the centralized computing node replies the update result to the edge computing node for different thresholds. Then, in the edge computing stage, the threshold parameter storage (storage format as Table 11) deployed on this node is modified for the threshold parameters that have been successfully modified in the reply result. Then, the edge computing node organizes the data that has not yet been sent to the centralized computing node, merges them according to the cell unit according to the method of 113), and outputs the data in the format of Table 7. Then, the edge computing node sends a test data append request to the centralized computing node, including the cell information corresponding to the data to be appended. Then, the centralized computing node returns a confirmation instruction to the edge computing node, including the cell information corresponding to the data to be appended. Then, the edge computing node sends the test data of each cell to the centralized computing node according to the cell dimension. Then, after receiving the test data of each cell, the centralized computing node processes and normalizes the test data according to the methods of 114) and 115), and appends the test data to the end of the existing test data set of the model.

[0388] Finally, the centralized computing node returns the update success information to the edge computing node for the cell data that has been processed. In this way, when the edge computing node executes 515) next time, for this cell, it only collects the data after the data has been successfully updated this time.

[0389] 516) The centralized computing node processes the training data set and the updated test data set according to the new marking threshold and the methods described in 117) to 119), and retrains the model, and sends the trained model parameters to the edge computing node.

[0390] First, the centralized computing node uses a sliding window to reorganize and mark the test data set according to the method described in 117). Then, the centralized computing node groups the test data set output in the previous step according to the method described in 118). Then, the centralized computing node uses the existing training data set and the test data set output in the previous step to train the model parameters according to the method described in 119). Finally, the centralized computing node sends the parameters of the model corresponding to the guarantee cell to the edge computing node according to the method described in 213).

[0391] In an embodiment of the present invention, the machine learning model parameters for dynamic adjustment of wireless network load are stored: the machine learning model used includes multiple layers, each layer contains parameters such as weights and biases, and is stored in a pth file (binary format). Furthermore, both the centralized computing node and the edge computing node store the pth files of the relevant model parameters and index them with a parameter model parameter database. The centralized computing node and the edge computing node also store an index table of model numbers and cell information. Such a data structure is conducive to the dynamic management of model parameters and supports the needs of multiple model parameter learning before and during the guarantee.

[0392] The embodiment of the present invention proposes a wireless network load data processing process adapted to a machine learning algorithm: the present invention collects the index tracking data of the centralized network management in the pre-training stage, including three tables of total throughput, number of users, and wireless resource usage. And read the three tables into the memory through the pandas library. And merge the three tables based on the "acquisition time". Further, the merged table fields are operated to generate new fields as the basis for marking and training, and the algorithm parameters need to be set according to the cell frequency band and bandwidth. In order to meet the requirements of machine learning, the relevant fields are normalized, and the specific parameters of the operation also need to be set according to the cell frequency band and bandwidth. The data of different cells and different time periods are grouped to form training data sets and test data sets of different models. In the above operation steps, the data processing, grouping and data set division methods all comprehensively consider the characteristics of the wireless network and the requirements of machine learning for the data format.

[0393] The embodiment of the present invention proposes a reorganization process of the training data set and the test data set: the embodiment of the present invention uses a sliding window to group the data set, the group size is 120 (2 minutes), the sliding step is 10 (10 seconds), and the final effect is to output a group of 12 1×120×7 tensors every 2 minutes. Such a data structure can be regarded as 12 120×7 photos, which meets the requirements of subsequent machine learning. The update cycle is 2 minutes, which ensures the real-time performance of data analysis and network parameter adjustment instructions.

[0394] The embodiment of the present invention proposes a labeling algorithm and mechanism for training data sets and test data sets: the present invention uniformly designs a set of labeling algorithms for training data sets and test data sets, and the algorithm includes a series of thresholds. By comparing the results of data element operations with the thresholds, the label value is finally output. The technical solution of the present invention defines the maximum value, minimum value and initial value of these thresholds, and the initial value of the threshold is used for the first labeling; these threshold values ​​are subsequently adjusted according to network changes to achieve the purpose of dynamically adjusting model parameters. Since there is a mapping relationship between labels and network adjustment instructions, such a design can achieve the mapping relationship between dynamically adjusting network load indicators and adjustment schemes.

[0395] The embodiment of the present invention proposes a more sophisticated model division: In order to make the machine learning model better reflect the wireless network situation, the present invention draws on the division principle of QCL (quasi-co-location) in the 5G wireless protocol, and sets a method for dividing each cell that needs to be protected in the same edge computing node into different models. First, the cell frequency band and bandwidth are divided to ensure that the frequency band and bandwidth of each cell in a model are the same. Secondly, further adjustments are made based on the results of model learning, and further divisions are made for models whose training data set accuracy or test data set accuracy does not meet the standards to ensure the accuracy of machine learning.

[0396] The embodiment of the present invention proposes a synchronization mechanism for model parameters: the present invention defines the storage scheme and data format of machine learning model parameters, as well as the mechanism for model parameter interaction between centralized computing nodes and edge computing nodes, and the mechanism for updating model parameters in centralized computing nodes and edge computing nodes. Through the application of these technologies, model parameter learning before assurance, dynamic model parameter learning during assurance, and dynamic application of model parameters are supported.

[0397] The embodiment of the present invention proposes a conversion mechanism from characteristic values ​​to base station parameter modification instructions: In the present invention, the calculated characteristic values ​​are first reversely restored to uplink characteristic values ​​and downlink characteristic values. Further, for each cell, according to the pre-set mapping rules, the current values ​​of various parameters are compared to minimize the generation of base station parameter modification instructions, and the signaling interaction between the edge computing node and the BBU is minimized while ensuring the accuracy of the instructions, thereby reducing the probability of errors.

[0398] The embodiment of the present invention proposes robust execution of base station parameter modification instructions: the present invention deploys an instruction execution (success) record table and an instruction execution failure record table, and establishes a data update mechanism for these two tables, so that the execution process of the base station parameter modification instructions has stronger traceability, facilitates rapid fault handling, and ensures the stability of system operation.

[0399] The embodiment of the present invention proposes a dynamic update mechanism for labeling parameters: The present invention realizes the dynamic update between base station load parameters and data features (labels) by establishing a dynamic update mechanism for threshold parameters defined in the data labeling algorithm and combining the deployment of machine learning algorithms. Combined with the conversion mechanism from feature values ​​to base station parameter modification instructions, the dynamic conversion of base station load parameters to base station parameter modification instructions is realized, adapting to the ever-changing wireless environment and ensuring a relatively good throughput rate when the base station load changes.

[0400] The embodiment of the present invention proposes a rule for merging multi-cell marking parameters: In view of the possible conflicts in marking parameter update requests issued by different cells in the same model at the same time and on the same edge computing node, the present invention designs a set of update request conflict resolution and merging rules based on the "voting" rule to improve the accuracy of marking parameter update requests.

[0401] The embodiment of the present invention proposes the introduction of a CCE utilization evaluation mechanism: based on the service channel load evaluation, the present invention adds the consideration of CCE utilization factors, reduces the probability of signaling congestion by increasing PDCCH resources when the CCE utilization rises, improves the stability of the load dynamic control mechanism, and improves user perception.

[0402] It should be noted that, for the method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that this implementation disclosure is not limited by the order of the actions described, because according to the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the present invention.

[0403] Fig.17 The block diagram of a wireless network load dynamic adjustment device provided by an embodiment of the present invention. The device is applied to an edge computing node, and includes: a parameter acquisition module 1701, a loading module 1702, a processing module 1703, a conversion module 1704, a first sending module 1705, an evaluation module 1706 and a second sending module 1707, wherein:

[0404] The parameter acquisition module 1701 is used to acquire the model parameters of the machine learning model of at least one cell in the protection area sent by the centralized computing node;

[0405] A loading module 1702 is used to load the model parameters to obtain a machine learning model of the corresponding cell, where the machine learning model is a model trained based on a multi-layer perception MLP neural network;

[0406] The processing module 1703 is used to use the machine learning model to process the traffic load data of the corresponding cell in real time to obtain a traffic decision point characteristic value;

[0407] A conversion module 1704, configured to convert the traffic decision point characteristic value into a base station parameter adjustment command;

[0408] A first sending module 1705 is configured to send the base station parameter adjustment command to a base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command;

[0409] An evaluation module 1706, configured to evaluate the real-time data obtained after adjusting the base station parameters and determine a threshold adjustment instruction;

[0410] The second sending module 1707 is used to send the threshold adjustment instruction to the centralized computing node, so that the centralized computing node performs corresponding threshold adjustment based on the threshold adjustment instruction, retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node.

[0411] Optionally, the acquisition module includes:

[0412] A first receiving module is used to receive a model parameter request sent by a centralized computing node, wherein the model parameter request includes: a model number and a corresponding cell;

[0413] A first search module, used to determine whether the cell in the model parameter request is found in the cell list associated with itself;

[0414] A third sending module, configured to send a confirmation message including the model number to the centralized computing node when the first searching module finds the cell in the model parameter request;

[0415] A second receiving module, configured to receive the model parameters sent by the centralized computing node, wherein the model parameters include the model number and the corresponding pth file;

[0416] The loading module is specifically used to load and store the pth file corresponding to the model number, obtain the machine learning model of the corresponding cell, and store the pth file of the model parameters.

[0417] Optionally, the processing module includes:

[0418] A data acquisition module, used to acquire the traffic load data of the corresponding cell, wherein the traffic load data includes: total throughput rate, number of users and wireless resource usage;

[0419] A merging module, used to merge the total throughput, the number of users and the wireless resource usage according to the collection time period of the corresponding cell;

[0420] A tensor determination module, used to determine the tensor of data points in the corresponding cell collection time period after merging;

[0421] The feature calculation module is used to use the model parameters of the corresponding cell to calculate the tensor and obtain the feature value of the traffic decision point.

[0422] Optionally, the conversion module includes:

[0423] A mode determination module, for determining the value range of the uplink and downlink modes of the corresponding cell respectively based on the characteristic value of the traffic decision point by using a sliding window for reorganization and marking inverse method;

[0424] A parameter setting module, configured to set the value of the network parameter according to the corresponding network parameter setting rule based on the determined value range of the uplink and downlink modes of the corresponding cell;

[0425] A mode acquisition module, used to obtain the values ​​of the uplink and downlink modes based on which the last command was issued in the command execution success record table of the corresponding cell;

[0426] The first comparison module is used to compare the values ​​of the uplink and downlink modes obtained by the mode acquisition module with the values ​​of the uplink and downlink modes determined by the mode determination module, and generate a parameter modification instruction Q for the corresponding cell according to the comparison result. t ;

[0427] The command conversion module is used to convert the parameter modification instruction of the corresponding cell into a base station parameter adjustment command that can be recognized by the BBU in the base station.

[0428] Optionally, the first comparison module includes:

[0429] A numerical comparison module, for comparing the numerical values ​​of the uplink and downlink modes based on the data with the determined numerical values ​​of the uplink and downlink modes;

[0430] The second query module is used to query the instruction execution success record table according to the comparison result;

[0431] A generation module is used to assign values ​​to the indexes of the key-value pairs of all column parameter names that meet the conditions queried by the second query module, and generate a parameter modification instruction Q for the corresponding cell. t .

[0432] Optionally, the command conversion module includes:

[0433] A parameter script generation module, used to generate a corresponding single-cell and single-parameter modification script based on each key-value pair of the parameter modification instruction of the corresponding cell;

[0434] A search module, configured to search for a mapping relationship between a parameter and a network management modification instruction parameter based on each key-value pair of the corresponding cell according to different data types of the parameter, and obtain a corresponding command name, a primary switch type IE and / or a secondary switch type IE;

[0435] The script conversion module is used to convert the corresponding command name, primary switch type IE and / or secondary switch type IE into a single parameter modification script of the corresponding cell, and use the single parameter modification script as a base station parameter adjustment command.

[0436] Optionally, the command conversion module further includes:

[0437] The parameter script merging module is used to merge the single parameter modification scripts according to the structure of the base station network management instructions to obtain a single-cell, multi-parameter modification script, wherein the two merged single parameter modification scripts belong to the same base station, the same BBU, and use the same command name.

[0438] Optionally, the evaluation module includes:

[0439] A first average value determination module, configured to determine the average uplink and downlink rates of each household in each cell according to a set time based on the real-time data obtained after adjusting the base station parameters;

[0440] A division module, used to divide the set time into a plurality of time periods, and calculate the values ​​of the uplink and downlink modes of the corresponding cell in each time period;

[0441] A frequency determination module, used to determine the values ​​of the uplink and downlink modes of each cell and the number of times the uplink and downlink modes appear according to each time period;

[0442] A majority merging module, used to merge the uplink and downlink majority of each cell and the corresponding number of occurrences;

[0443] A second average value determination module is used to query the average uplink and downlink rate of each household in the corresponding cell based on the uplink and downlink modes and the corresponding number of occurrences of each cell after merging;

[0444] The average value combination module is used to combine the average uplink and downlink rates of each household in all cells to obtain the corresponding array of average uplink and downlink rates of each household.

[0445] A second comparison module is used to compare the value of each element in the array of average uplink and downlink rates per household with the corresponding target value;

[0446] The threshold adjustment generation module is used to generate the corresponding and T upllnk，prb Adjustment instructions for three thresholds.

[0447] Optionally, the device further comprises:

[0448] An indicator acquisition module, used for acquiring a current control channel element CCE utilization index of each cell after the first sending module sends the base station parameter adjustment command to the base station;

[0449] A load determination module, configured to determine a current CCE load of a corresponding cell based on the CCE utilization index;

[0450] A channel occupancy adjustment instruction generating module, used to generate a channel occupancy adjustment instruction when the current CCE load of the cell is greater than a set threshold;

[0451] The fourth sending module is used to send the channel occupancy adjustment instruction to the base station, so that the base station adjusts the channel parameters based on the channel occupancy adjustment instruction.

[0452] Optionally, the load determination module includes:

[0453] A utilization determination module, used to determine the total CCE utilization of each cell at each moment based on the CCE utilization index;

[0454] The average utilization rate determination module is used to determine the average CCE utilization rate of each cell within a set time period based on the total CCE utilization rate of each cell, and use the average CCE utilization rate as the current CCE load status of the corresponding cell.

[0455] See also Fig.18 , a block diagram of a training device for a machine learning model is provided in an embodiment of the present invention, the device is applied to a centralized computing node, and includes: an initialization module 1801, an acquisition module 1802, a training module 1803 and a determination module 1804, wherein,

[0456] Initialization module 1801, used to initialize model parameters of the machine learning model;

[0457] An acquisition module 1802 is used to acquire a training data set and a test data set;

[0458] The training module 1803 is used to perform multiple rounds (each round is an epoch) of training on the machine learning model after initializing the model parameters based on the training data set; wherein each round of training includes: initializing the measurement variables of the training data set, reading the training data set and its labels, and storing the label values ​​of the labels; traversing the training data set, during the traversal process, calculating the characteristic values ​​using the model parameters, and storing the calculated characteristic values ​​as the characteristic values ​​output by this batch of training data, using the loss function to calculate the loss value, calculating the gradient based on the loss value, performing back propagation, and updating the machine learning model function, updating the measurement variables, and storing the loss value as the calculation variable; reading the stored label values ​​and the characteristic values ​​output by this batch of training data, calculating the accurate number of this batch of training data based on the label values ​​and the characteristic values, and updating the accurate number and the total number to the measurement variables, and superimposing the original values ​​to obtain the calculation variables including the accurate number, the total number and the loss value; repeating the above process until all the theoretical training is completed, and outputting the accuracy of the corresponding training results;

[0459] The determination module 1804 is used to determine that the machine learning model is a trained model when the accuracy reaches a threshold.

[0460] Optionally, the acquisition module includes:

[0461] A data acquisition module is used to obtain the traffic data of each cell in the high-load protection area;

[0462] A merging module, used to merge the acquired traffic data in units of cells to obtain a cell-level merging table;

[0463] The field generation module is used to operate on the fields of the cell-level merge table and generate new fields as the basis for labeling and training;

[0464] A normalization processing module, used for normalizing rate-related fields according to the network configuration of each cell;

[0465] The grouping module is used to group the data of different time periods of each cell after normalization processing to obtain a training data set and a test data set.

[0466] Optionally, the acquisition module further includes:

[0467] The reorganization and labeling module is used to reorganize and label the training data set and the test data set respectively using a sliding window.

[0468] Optionally, an embodiment of the present invention further provides an electronic device, including:

[0469] processor;

[0470] a memory for storing instructions executable by the processor;

[0471] Wherein, the processor is configured to execute the instructions to implement the wireless network load dynamic adjustment method as described above or the machine learning model training method as described above.

[0472] Optionally, an embodiment of the present invention also provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the wireless network load dynamic adjustment method as described above or the machine learning model training method as described above.

[0473] Optionally, an embodiment of the present invention also provides a computer program product, including a computer program or instructions, which, when executed by a processor of an electronic device, implements the wireless network load dynamic adjustment method as described above or the machine learning model training method as described above.

[0474] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0475] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0476] Fig.19 It is a block diagram of an electronic device 1900 provided by an embodiment of the present invention. As shown in the figure, it includes a processor 1901, a communication interface 1902, a memory 1903 and a communication bus 1904, wherein the processor 1901, the communication interface 1902, and the memory 1903 communicate with each other through the communication bus 1904;

[0477] Memory 1903, used to store the processor executable instructions;

[0478] The processor 1901 is used to implement the above method when executing the executable instructions in the memory 1903.

[0479] The communication bus in this embodiment may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0480] The communication interface is used for communication between the above electronic device and other devices.

[0481] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0482] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0483] In another embodiment provided by the present invention, a computer-readable storage medium is also provided, and when the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device can execute the wireless network load dynamic adjustment method shown above or the training method of the machine learning model as described above. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0484] In another embodiment provided by the present invention, a computer program product is also provided, including a computer program or instructions. When the computer program or instructions are executed by a processor of an electronic device, the electronic device executes the wireless network load dynamic adjustment method as described above or the machine learning model training method as described above.

[0485] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0486] Fig. 20 2 is a block diagram of an apparatus 2000 for dynamically adjusting wireless network load or training a machine learning model provided by an embodiment of the present invention. For example, the apparatus 2000 may be provided as a server. Fig. 20 , the apparatus 2000 includes a processing component 2022, which further includes one or more processors, and a memory resource represented by a memory 2032 for storing instructions, such as an application, that can be executed by the processing component 2022. The application stored in the memory 2032 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 2022 is configured to execute the instructions to perform the above method.

[0487] The device 2000 may also include a power supply component 2026 configured to perform power management of the device 2000, a wired or wireless network interface 2050 configured to connect the device 2000 to a network, and an input / output (I / O) interface 2058. The device 2000 may operate based on an operating system stored in the memory 2032, such as Windows Server™, MacOS X™, Unix™, Linux™, FreeBSD™, or the like.

[0488] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.

[0489] It should be understood that the present invention is not limited to the exact construction that has been described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.< / switch> < / key2> < / instruction> < / key> < / instruction> < / key> < / instruction> < / key> < / instruction> < / key> < / instruction> < / switch> < / key> < / instruction> < / key1> < / valuei1> < / key1>

Claims

1. A method for dynamically adjusting wireless network load, characterized in that: The method is applied to an edge computing node, comprising: Obtain model parameters of a machine learning model of at least one cell in a security area sent by a centralized computing node; Load the model parameters to obtain a machine learning model for the corresponding cell, where the machine learning model is a model trained based on a multi-layer perception MLP neural network; Using the machine learning model to process the traffic load data of the corresponding cell in real time to obtain a traffic decision point characteristic value; Converting the traffic decision point characteristic value into a base station parameter adjustment command; Sending the base station parameter adjustment command to a base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command; Evaluating the real-time data obtained after adjusting the base station parameters to determine a threshold adjustment instruction; The threshold adjustment instruction is sent to the centralized computing node, so that the centralized computing node performs corresponding threshold adjustment based on the threshold adjustment instruction, retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node.

2. The method for dynamically adjusting wireless network load according to claim 1, characterized in that: The obtaining of model parameters of the machine learning model of at least one cell in the security area sent by the centralized computing node includes: Receiving a model parameter request sent by a centralized computing node, the model parameter request including: a model number and a corresponding cell; When the cell in the model parameter request is found in the cell list associated with itself, sending a confirmation message including the model number to the central computing node; Receiving model parameters sent by the centralized computing node, the model parameters including: a model number and a corresponding pth file; The loading of the model parameters to obtain the machine learning model of the corresponding cell specifically includes: loading the pth file corresponding to the model number, obtaining the machine learning model of the corresponding cell, and storing the pth file of the model parameters.

3. The method for dynamically adjusting wireless network load according to claim 1, characterized in that: The using the machine learning model to process the traffic load data of the corresponding cell in real time to obtain the traffic decision point characteristic value includes: Acquire traffic load data of the corresponding cell, the traffic load data including: total throughput rate, number of users and wireless resource usage; Combining the total throughput rate, the number of users and the wireless resource usage according to the collection time period of the corresponding cell; Determine the tensor of data points within the corresponding cell collection time period after merging; The tensor is operated using the model parameters of the corresponding cell to obtain the characteristic value of the traffic decision point.

4. The method for dynamically adjusting wireless network load according to claim 1, characterized in that: Converting the traffic decision point characteristic value into a base station parameter adjustment command comprises: Based on the traffic decision point characteristic value, a sliding window is used to perform reorganization and marking inverse method to respectively determine the value range of the uplink and downlink modes of the corresponding cell; Based on the determined value ranges of the uplink and downlink modes of the corresponding cell, setting the value of the network parameter according to the corresponding network parameter setting rule; Obtain the values ​​of the uplink and downlink modes based on which the last command was issued in the command execution success record table of the corresponding cell; Comparing the values ​​of the uplink and downlink modes based on the data with the determined values ​​of the uplink and downlink modes, and generating a parameter modification instruction for the corresponding cell according to the comparison result; The parameter modification instruction of the corresponding cell is converted into a base station parameter adjustment command that can be recognized by the BBU in the base station.

5. The method for dynamically adjusting wireless network load according to claim 4, characterized in that: The step of comparing the values ​​of the uplink and downlink modes according to which the uplink and downlink modes are used with the determined values ​​of the uplink and downlink modes, and generating a parameter modification instruction for the corresponding cell according to the comparison result, comprises: Comparing the values ​​of the uplink and downlink modes based on the data with the determined values ​​of the uplink and downlink modes; According to the comparison result, the query command execution success record table is queried, the indexes of the key-value pairs of all column parameter names that meet the conditions are assigned, and the parameter modification instructions of the corresponding cell are generated.

6. The method for dynamically adjusting wireless network load according to claim 4, characterized in that: The converting the parameter modification instruction of the corresponding cell into a base station parameter adjustment command recognizable by a BBU in the base station includes: Based on each key-value pair of the parameter modification instruction of the corresponding cell, generating a corresponding single-cell and single-parameter modification script; According to different data types of the parameters, based on each key-value pair of the corresponding cell, searching for a mapping relationship between the parameters and the network management modification instruction parameters, obtaining a corresponding command name, a primary switch type IE and / or a secondary switch type IE; The corresponding command name, primary switch type IE and / or secondary switch type IE are converted into a single parameter modification script of the corresponding cell, and the single parameter modification script is used as a base station parameter adjustment command.

7. The method for dynamically adjusting wireless network load according to claim 6, characterized in that: The converting the parameter modification instruction into a base station parameter adjustment command recognizable by the BBU in the base station also includes: According to the structure of the base station network management instruction, the single parameter modification scripts are merged to obtain a single cell, multi-parameter modification script, wherein the two merged single parameter modification scripts belong to the same base station, the same BBU, and use the same command name.

8. The method for dynamically adjusting wireless network load according to claim 1, characterized in that: The step of evaluating the real-time data obtained after adjusting the base station parameters and determining the threshold adjustment instruction includes: For the real-time data obtained after adjusting the base station parameters, determine the average uplink and downlink rates of each household in each cell according to the set time; Divide the set time into multiple time periods, and calculate the values ​​of the uplink and downlink modes of the corresponding cell in each time period; Determine the values ​​of uplink and downlink modes of each cell and the number of occurrences of the uplink and downlink modes according to each time period; Merge the uplink and downlink modes of each cell and the corresponding number of occurrences respectively; Based on the merged uplink and downlink mode of each cell and the corresponding number of occurrences, query the average uplink and downlink rate of each household in the corresponding cell; The average uplink and downlink rates of each household in all cells are combined to obtain the corresponding array of average uplink and downlink rates of each household; The value of each element in the array of average uplink and downlink rates for each household is compared with the corresponding target value, and an adjustment instruction for the corresponding threshold is generated.

9. The method for dynamically adjusting wireless network load according to any one of claims 1 to 8, characterized in that: After sending the base station parameter adjustment command to the base station, the method further includes: Obtain the current control channel unit CCE utilization index of each cell; Determine a current CCE load of the corresponding cell based on the CCE utilization index; When the current CCE load of the cell is greater than a set threshold, generating a channel occupancy adjustment instruction; The channel occupancy adjustment instruction is sent to the base station, so that the base station adjusts the channel parameters based on the channel occupancy adjustment instruction.

10. The method for dynamically adjusting wireless network load according to claim 9, characterized in that: The determining the current CCE load of the corresponding cell based on the CCE utilization index includes: Determine the total CCE utilization of each cell at each moment based on the CCE utilization index; The average CCE utilization rate of each cell within a set time period is determined based on the total CCE utilization rate of each cell, and the average CCE utilization rate is used as the current CCE load condition of the corresponding cell.

11. A method for training a machine learning model, characterized in that: The method is applied to a centralized computing node, comprising: Initialize the model parameters of the machine learning model; Get training data set and test data set; Based on the training data set, the machine learning model after initializing the model parameters is trained for multiple rounds, wherein each round of training includes: initializing the measurement variables of the training data set, reading the training data set and its label, and storing the label value of the label; traversing the training data set, during the traversal process, calculating the characteristic value using the model parameters, and storing the calculated characteristic value as the characteristic value output of this batch of training data, using the loss function to calculate the loss value, calculating the gradient based on the loss value, performing back propagation, and updating the machine learning model function, updating the measurement variables, and storing the loss value as the calculation variable; reading the stored label value and the characteristic value output of this batch of training data, calculating the accurate number of this batch of training data based on the label value and the characteristic value, and updating the accurate number and the total number to the measurement variable, and superimposing the original value to obtain the calculation variable including the accurate number, the total number and the loss value; repeating the above process until all the theoretical training is completed, and outputting the accuracy of the corresponding training result; When the accuracy reaches a threshold, the machine learning model is determined to be a trained model.

12. The method for training a machine learning model according to claim 11, characterized in that: The obtaining of the training data set and the test data set includes: Obtain the traffic data of each cell in the high-load protection area; The acquired traffic data is merged according to the cell level to obtain a cell-level merge table; Operate on the fields of the cell-level merge table to generate new fields as the basis for labeling and training; Normalize rate-related fields according to the network configuration of each cell; The normalized data of different time periods in each cell are grouped to obtain training data sets and test data sets.

13. The method for training a machine learning model according to claim 11, characterized in that: The obtaining of the training data set and the test data set also includes: The training data set and the test data set are reorganized and labeled using sliding windows respectively.

14. A wireless network load dynamic adjustment device, characterized in that: The device is applied to an edge computing node, including: A parameter acquisition module, used to acquire model parameters of a machine learning model of at least one cell in a security area sent by a centralized computing node; A loading module, used to load the model parameters to obtain a machine learning model of the corresponding cell, wherein the machine learning model is a model obtained by training based on a multi-layer perception MLP neural network; A processing module, used to use the machine learning model to process the traffic load data of the corresponding cell in real time to obtain a traffic decision point characteristic value; A conversion module, used for converting the traffic decision point characteristic value into a base station parameter adjustment command; A first sending module, configured to send the base station parameter adjustment command to a base station, so that the base station adjusts the corresponding base station parameters according to the base station parameter adjustment command; An evaluation module, used to evaluate the real-time data obtained after adjusting the base station parameters and determine a threshold adjustment instruction; The second sending module is used to send the threshold adjustment instruction to the centralized computing node, so that the centralized computing node performs corresponding threshold adjustment based on the threshold adjustment instruction, retrains the machine learning model, and sends the model parameters of the retrained machine learning model to the edge computing node.

15. A training device for a machine learning model, characterized in that: The device is applied to a centralized computing node, and includes: Initialization module, used to initialize the model parameters of the machine learning model; An acquisition module is used to obtain training data sets and test data sets; A training module is used to perform multiple rounds of training on the machine learning model after initializing the model parameters based on the training data set; wherein each round of training includes: initializing the measurement variables of the training data set, reading the training data set and its labels, and storing the label values ​​of the labels; traversing the training data set, during the traversal process, calculating the characteristic values ​​using the model parameters, and storing the calculated characteristic values ​​as the characteristic values ​​output by this batch of training data, using the loss function to calculate the loss value, calculating the gradient based on the loss value, performing back propagation, and updating the machine learning model function, updating the measurement variables, and storing the loss value as a calculation variable; reading the stored label values ​​and the characteristic values ​​output by this batch of training data, calculating the accurate number of this batch of training data based on the label values ​​and the characteristic values, and updating the accurate number and the total number into the measurement variables, and superimposing the original values ​​to obtain a calculation variable including the accurate number, the total number and the loss value; repeating the above process until all the theoretical training is completed, and outputting the accuracy of the corresponding training results; A determination module is used to determine that the machine learning model is a trained model when the accuracy reaches a threshold.

16. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the method for dynamically adjusting the wireless network load as described in any one of claims 1 to 10 or the method for training the machine learning model as described in any one of claims 11 to 13.

Images