A method and device for elastic scaling of slices based on network perception

Abstract

The present invention discloses a method and apparatus for elastically scaling slices based on network perception. The method includes: collecting service experience KQI indicators in real time; determining whether the network is normal based on the collected service experience KQI indicators; when the network is normal, determining whether the slice needs to be scaled based on the collected service experience KQI indicators and knowledge from an expert knowledge base; and performing scaling operations on the slice if it is determined that scaling is required.

Description

Technical Field

[0001] The present invention relates to elastic scaling in 5G slicing, and in particular to a method and device for elastic scaling of slices based on network perception. Background Art

[0002] The 5G slice end-to-end management system is required to support closed-loop fault recovery. Typical slice management systems are divided into design and execution domains. On the design side, closed-loop fault recovery involves defining policies for the VNFs (Virtualized Network Functions) of the relevant slices / subslices. Actions are then formulated based on performance SLAs and related alarms. In the operational state, actions are executed according to the policies based on the operating environment.

[0003] Take, for example, the implementation of slice elastic scaling for VR live streaming. VR (virtual reality) live streaming is a high-bandwidth service, a slice type of the eMBB (Enhanced Mobile Broadband) foundational service, which includes radio / bearing / core network subslices. For example, the core network subslice UPF (User Plane Function) handles data uplink and downlink, packet routing, and forwarding. As service data volume increases, the UPF load increases. (Note that due to the wide variety of slice / subslice services, the following examples use the UPF of the eMBB slice and core network subslice for VR live streaming.)

[0004] On the design side, corresponding SLA indicators are defined, such as a throughput threshold upper limit of 1024 Mbps, and related closed-loop self-healing and scaling policies are defined. If the throughput exceeds the upper threshold for 60 seconds, an action is triggered, which is to expand the UPF capacity and add a VNFC (VNFComponent).

[0005] In the running state, through the collection, monitoring and corresponding intelligent analysis of data indicator SLA and alarm data, it is known that the throughput exceeds the limit. According to the corresponding policy action UPF defined by the design side, the scale-up command is executed to scale up and down VNFC, which actually adds a VM virtual machine to share the corresponding load, thus achieving the corresponding self-healing effect.

[0006] The bearer network sub-slice SMF also has corresponding bandwidth expansion design and policy actions, and its design and closed-loop action principles are the same.

[0007] If the throughput threshold continues to reach the new upper limit for a period of time after capacity expansion, capacity will be expanded again until it stabilizes.

[0008] This design and system is a common practice for end-to-end elastic scaling and closed-loop self-healing in 5G slices. However, there are three scenarios that can affect user experience:

[0009] During the period after throughput reaches the threshold but before scaling down (e.g., 60 seconds), user perception is degraded.

[0010] If capacity is still far from meeting throughput requirements after scaling up or down, it will be necessary to wait for a while for trigger conditions to trigger further capacity expansion. During this time, user experience will continue to deteriorate.

[0011] If the throughput reaches the upper threshold but does not exceed the specified time, then decreases, but then rebounds with short, erratic intervals, this is an oscillating, pulse-like trend. In this scenario, capacity expansion is impossible or targeted capacity expansion is difficult to design, and user experience deteriorates.

[0012] In these scenarios, if it's impossible to lower the threshold, a common solution is to do it all at once very quickly after reaching the threshold. However, if the data only surges momentarily and then immediately falls back, it will waste slice resources and it will be impossible to determine under what conditions the data will rebound and recover. Similarly, simply lowering the threshold is not a scientific approach.

[0013] From the above analysis, we can see that the simple threshold policy design method has certain defects. It is a static threshold policy design, which is relatively rigid and difficult to calculate the optimal value policy, and cannot effectively improve user perception.

[0014] The current industry network perception indicators for cloud VR and cloud gaming include: the device-side (5G CPE) service experience KQI indicator, the 5G slice service quality KQI, and the 5G slice network performance KPI. The specific indicators are as follows:

[0015] Service experience KQI indicators include interactive experience indicators (initial buffering success rate, initial buffering delay, MTP delay + black edge rate (an indicator for reducing VR sickness)); media quality (resolution, bit rate, frame rate); and continuity indicators (stuttering duration, screen distortion and packet loss rate).

[0016] · Corresponding 5G slice service quality KQI includes xKB startup delay, video download rate, video rate bit rate ratio, MTP pipe side delay

[0017] 5G slicing network performance KPIs include throughput, latency, jitter, rate, number of online users, etc.

[0018] Currently, operators have designed relevant thresholds for these indicators. However, due to the large number of indicators, business-oriented operations and maintenance are overly complex and redundant, and the degree of automation is relatively low.

[0019] In the future, with the promotion and application development of 5G, eMBB, Urllc (ultra reliable and low latency communication), and mMTC (massive machine type of communication) hybrid slicing services will also be launched, and the complexity of services and applications will follow. Therefore, the rational use of slice resources for scaling and ensuring network awareness will be an important principle of 5G slice end-to-end management. Summary of the Invention

[0020] The purpose of the present invention is to provide a method and device for slice elastic scaling based on network perception, which can respond quickly in real-time operation and improve user perception.

[0021] According to a first aspect of the present invention, a method for slice elastic scaling based on network awareness is provided, comprising:

[0022] Real-time collection of business experience KQI indicators;

[0023] Determine whether the network is normal based on the collected service experience KQI indicators;

[0024] When the network is judged to be normal, it is determined whether the slice needs to be scaled based on the collected service experience KQI indicators and the knowledge of the expert knowledge base, and if it is determined that the slice needs to be scaled, the scaling operation is performed on the slice.

[0025] Preferably, the real-time collection of service experience KQI indicators includes: real-time collection of slice service quality KQI and slice network performance KPI.

[0026] Preferably, the scaling operation on the slices includes: determining the number of slices that need to be scaled using a slice scaling model and real-time network data; and increasing or decreasing slices according to the determined number of slices that need to be scaled.

[0027] Preferably, the network real-time data includes: KPI indicators including traffic and CPU memory and real-time business data.

[0028] Preferably, the slice scaling model is obtained by machine learning or deep learning of historical data, including converting historical data into feature data and then performing machine learning or deep learning.

[0029] Preferably, the slice scaling model is a linear regression model on slice scaling.

[0030] Preferably, the determining the number of slices to be scaled using the slice scaling model and network characteristic data includes:

[0031] After converting the network real-time data into feature data, substituting the data into a linear regression model to obtain a first value;

[0032] Dividing the first value by the initial value when not stretched to obtain a second value;

[0033] Dividing the second value by n, rounding up the result to obtain a required scaling factor, and then multiplying the required scaling factor by the current slice data to obtain the slice data to be scaled;

[0034] Wherein, n is a positive integer.

[0035] According to a second aspect of the present invention, a device for slice elastic scaling based on network awareness is provided, comprising:

[0036] Collection module, used to collect business experience KQI indicators in real time;

[0037] The network fault judgment module is used to judge whether the network is normal based on the collected service experience KQI indicators;

[0038] The slice scaling determination module is used to determine whether the slice needs to be scaled based on the collected service experience KQI indicators and the knowledge of the expert knowledge base when it is judged that the network is normal, and to perform scaling operations on the slice when it is determined that the slice needs to be scaled.

[0039] Preferably, the real-time collection of service experience KQI indicators includes: real-time collection of slice service quality KQI and slice network performance KPI.

[0040] Preferably, the slice scaling determination module includes:

[0041] Slice scaling model, used to determine the number of slices that need to be scaled based on real-time network data;

[0042] The scaling execution unit is configured to increase or decrease slices according to the determined number of slices that need to be scaled.

[0043] Compared with existing technologies, the present invention has the following advantages: 1) elastically scalable classification and regression models are generated through machine learning, enabling rapid response in real-time operation and improving user experience; 2) by separating the design domain from the execution domain, the entire execution domain process, except for parameter adjustment, is fully automated and flexible.

[0044] The present invention will be described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1A Schematic diagram of a network-aware slice elastic scaling method of the present invention;

[0046] Figure 1B This is a schematic diagram of a network-aware slice elastic scaling device of the present invention.

[0047] Figure 2 This is a schematic diagram of the entire system architecture and process of the method for implementing network-aware slice elastic scaling of the present invention;

[0048] Figure 3 It is a dynamic elasticity and contraction schematic diagram of the present invention;

[0049] Figure 4 This is a schematic diagram of calculating the adaptive scaling index of eMBB slices according to the present invention;

[0050] Figure 5 is a schematic diagram of the execution engine of the present invention;

[0051] Figure 6 It is a schematic diagram of the feature definition and feature engineering process of the present invention. DETAILED DESCRIPTION

[0052] Figure 1A The present invention shows a network-aware slice elastic scaling method, such as Figure 1A As shown, the method of the present invention includes: collecting service experience KQI indicators in real time; judging whether the network is normal based on the collected service experience KQI indicators; when it is judged that the network is normal, determining whether the slice needs to be scaled based on the collected service experience KQI indicators and the knowledge of the expert knowledge base, and performing scaling operations on the slice when it is determined that the slice needs to be scaled.

[0053] In one example of the present invention, real-time collection of service experience KQI indicators may include: real-time collection of slice service quality KQI and slice network performance KPI.

[0054] In one embodiment of the present invention, scaling a slice includes determining the number of slices to be scaled using a slice scaling model and real-time network data; and increasing or decreasing slices based on the determined number of slices to be scaled. The real-time network data includes KPIs for traffic, CPU, and memory, as well as real-time service data. The slice scaling model is derived through machine learning or deep learning of historical data, including converting historical data into feature data and then performing machine learning or deep learning.

[0055] In one embodiment of the present invention, the slice scaling model is a linear regression model for slice scaling. Alternatively, the slice scaling model can also be various neural networks.

[0056] In a preferred embodiment of the present invention, determining the number of slices that need to be scaled using a slice scaling model and network feature data includes: converting the real-time network data into feature data and substituting the data into a linear regression model to obtain a first value; dividing the first value by an initial value when not scaled to obtain a second value; dividing the second value by n and rounding up to obtain a required scaling factor, i.e., an adaptive scaling index of the eMBB slice; and multiplying the required scaling factor by the current slice data to obtain the slice data that needs to be scaled; wherein n is a positive integer.

[0057] Figure 1B The present invention shows a network-aware slice elastic scaling device, such as Figure 1B As shown, the device of the present invention includes: a collection module for collecting service experience KQI indicators in real time; a network fault judgment module for judging whether the network is normal based on the collected service experience KQI indicators; a slice scaling determination module for determining whether the slice needs to be scaled based on the collected service experience KQI indicators and the knowledge of the expert knowledge base when it is judged that the network is normal, and performing scaling operations on the slice when it is determined that the slice needs to be scaled.

[0058] As described above, the real-time collection of service experience KQI indicators includes: real-time collection of slice service quality KQI and slice network performance KPI.

[0059] In one example of the present invention, the slice scaling determination module includes: a slice scaling model, which is used to determine the number of slices that need to be scaled based on real-time network data; and a scaling execution unit, which is used to increase or decrease slices based on the determined number of slices that need to be scaled.

[0060] The basic idea of ​​this invention is to use indicator data analysis and machine learning to obtain the final slice scaling model, that is, first define the characteristics of the slice, then mark and analyze the cleaned data, and then use machine learning to classify it. By analyzing the classification results, strategy formulation is carried out to obtain the final strategy model.

[0061] The present invention can be divided into Figure 2 The design domain and execution domain are shown. Data design is performed in the design domain. In the execution domain, the machine learning model and optimization algorithm are specified, machine learning effect evaluation is defined, parameters are adjusted for poor learning results, and strategy design is implemented after achieving the desired learning results. On the execution side, data preparation (data acquisition and transformation) is performed, along with feature engineering, machine learning / deep learning, evaluation and parameter adjustment to obtain a model, model conversion to a strategy, and strategy execution.

[0062] The following combination Figures 2 to 5 The specific implementation steps of the present invention are described in detail.

[0063] The first step is data design.

[0064] On the design side, design which KQI / KPI indicators need to be collected and design the data for machine learning, mainly data definition and feature engineering.

[0065] Currently, the industry's network perception indicators for cloud VR and cloud gaming include the device-side (5G CPE) service experience KQI, the 5G slice service quality KQI, and the 5G slice network performance KPI. The specific indicators are as follows:

[0066] Service experience KQI indicators include interactive experience indicators (initial buffering success rate, initial buffering delay, MTP delay + black edge rate (an indicator for reducing VR sickness)); media quality (resolution, bit rate, frame rate); and continuity indicators (stuttering duration, screen distortion and packet loss rate).

[0067] · Corresponding 5G slice service quality KQI includes xKB startup delay, video download rate, video rate bit rate ratio, MTP pipe side delay

[0068] 5G slicing network performance KPIs, such as throughput, latency, jitter, rate, and number of online users

[0069] These indicators are not independent. It can be simply understood that the service experience KQI indicator is determined by the 5G slice service quality KQI, and the 5G slice service quality KQI is determined by the lower-level 5G slice network performance KPI.

[0070] It can also be understood in this way: when the user experience is relatively poor, the 5G slice service quality KQI and 5G slice network performance KPI can be used to determine whether there is an equipment failure on the 5G end side or in the slice pipeline. If not, then the user physical examination, that is, the service experience KQI indicator, can be used directly to determine whether scaling up or down is needed based on the knowledge base in the knowledge graph.

[0071] Whether to scale back or expand is determined directly based on the expert experience in the knowledge graph, rather than using machine learning or deep learning. This is because it can almost be judged by the user-oriented business experience KQI indicators. More importantly, machine learning or deep learning obtains a probability. If the probability is around 50%, or if a low-probability event occurs, and the probability is too low to scale back, then a relatively large perception problem will arise.

[0072] First determine whether scaling is necessary, and then determine the scale of scaling. Even if the scale is not enough, the perception problem caused by insufficient scaling can be avoided to a certain extent.

[0073] If scaling is determined to be necessary, feature engineering is performed on real-time network data, which converts real-time network data into feature data, and the final scaling scale is obtained through machine learning / deep learning.

[0074] Different slices, sub-slices, and related VNFs manifest themselves differently in their respective service processing, and their abstracted features are also different. For example, although VR live streaming and VR gaming both belong to eMBB basic services, they are different types of slices. For example, their latency requirements are different. VR gaming obviously requires higher latency, while VR live streaming has relatively loose latency requirements.

[0075] Taking the UPF scaling of core network slices associated with VR live streaming as an example, the following data is required as the final relevant features:

[0076] Slicing basic services, such as eMBB basic services;

[0077] Slice type, such as VR live broadcast;

[0078] Slice network performance KPIs, such as throughput, latency, jitter, and rate;

[0079] KPI data such as CPU and memory indicators of the VNFC virtual machine.

[0080] Once the data is defined, it is also necessary to specify how it should be vectorized and normalized.

[0081] Vectorization: Each feature can be decomposed into one or more dimensions. For example, the memory indicator can be designed as 512M / 1G / 2G / 4G / 8G.

[0082] Data standardization, also known as normalization in machine learning.

[0083] There are three standardized designs. One is to design the data as 0 / 1 or -1 / 1. For example, if the design is 0 / 1, taking the memory indicator above as an example, assuming that the memory occupies 1.2G, then 2G is 1, and the other items are 0;

[0084] One is to subtract the mean from the data and then divide it by the variance, such as throughput;

[0085] Another thing is to design which data are outliers.

[0086] In addition to data definition, data design also requires specifying the ratio of training data and test data.

[0087] The second step is to specify the machine learning operator.

[0088] Select the operator to be learned (in some systems or documents, such as Li Hang's "Statistical Learning Methods", training algorithms such as logistic regression, SVM, and KNN are called models, and optimization algorithms for minimizing loss functions such as gradient descent and Newton iteration are called algorithms. However, in Alibaba and our framework, training algorithms, loss functions, optimization algorithms, and regularization are collectively referred to as operators) and hyperparameters.

[0089] Since it is desired to respond as quickly as possible without affecting user perception, the elasticity system of the present invention uses a regression model to directly elasticize in one step.

[0090] First, choose whether to use machine learning or deep learning. This depends on the amount of data and the current learning environment, such as computing resources.

[0091] If the amount of data is not large, choosing deep learning is very likely to cause overfitting, and it is also more troublesome to adjust the hyperparameters, so the following explanation will take machine learning as an example.

[0092] Regression learning can be used to determine the degree of elasticity and shrinkage, such as whether to add one or two VNFC virtual machines. Standard linear regression can be used.

[0093] The optimization algorithm, that is, whether to use stochastic gradient descent, mini-batch gradient descent, momentum gradient descent algorithm, or AdaGrad, etc., has little difference in efficiency in the slice shrinking device.

[0094] Hyperparameters, such as the number of hidden layers in a neural network.

[0095] Step 3: Definition of evaluation of learning outcomes.

[0096] Regression learning uses general error algorithms such as mean absolute error (MAE) and mean squared error (MSE). Similarly, specific errors need to be defined.

[0097] Mainly, if the machine learning effect evaluation does not meet the target, the parameters need to be readjusted.

[0098] There are several types of parameter adjustments:

[0099] 1. Redesign the data, for example, if the training data is too small, it is easy to overfit;

[0100] 2. Re-select features to see if some features are not taken into consideration;

[0101] 3. Redefine regularization, especially when the training data performs well but the test data performs poorly.

[0102] After the settings are reset, the machine learning will be re-performed according to the new parameters in the running state to finally output the model.

[0103] This model is the model learned by the machine, that is, the designation of certain operators and these features (specific dimensions) and their respective weights. The whole is a model.

[0104] Step 4: Indicator design and strategy definition.

[0105] It is necessary to design adaptive scaling indicators for eMBB slices.

[0106] Note that these two indicators cannot be directly obtained through linear or nonlinear calculations (such as kernel functions). They need to be calculated by obtaining feature weights through machine learning.

[0107] refer to Figure 4 Formula 1 is the initial value without shrinkage, and Formula 2 represents the linear regression model, where θ is the weight of each feature dimension, which needs to be learned by machine learning.

[0108] Formula 3 represents the regression calculation divided by the initial value. Formula 4 gives the final calculation of the adaptive elasticity index. Finally, rand is rounded to get the specific elasticity.

[0109] After the model is generated, a strategy needs to be designed. The current strategy is relatively simple and can be designed for the following situations:

[0110] 1. Calculate whether the eMBB slice network perception degradation rate indicator needs to be scaled up or down based on the model;

[0111] 2. Calculate whether the eMBB slice network perception degradation rate indicator needs to be restored based on the model;

[0112] 3. Calculate the adaptive scaling index of the eMBB slice based on the model to obtain the scaling parameters;

[0113] For example, VR's UPF strategy is as follows:

[0114] Policy: UPF expands two instances

[0115] Action: 1. Obtain VNFC based on resources. 2. Add a VM to each instance.

[0116] Step 5: Process definition.

[0117] Define the automation processes and inputs and outputs for all execution domains.

[0118] The major process is data preparation / feature engineering / classification model learning / evaluation (parameter adjustment) / regression model learning / evaluation (parameter adjustment) / model release / production strategy model.

[0119] Each large process has its own internal processes, which will be introduced in the execution domain below.

[0120] Except for the first and last steps, the input of each step is the output of the previous step, and the output of each step is the input of the next step.

[0121] The above steps 1 to 5 are all carried out on the design side, and the following step 6 and onwards are the execution process in the running state.

[0122] Step 6: Data preparation.

[0123] The specific steps of data preparation are collecting data, cleaning data, standardizing data, labeling data, and splitting training data and test data.

[0124] Data collection refers to collecting historical data of slices and sub-slices designed by the design end, including KPI indicators such as traffic, CPU memory, and current business data.

[0125] Clean the data and remove abnormal data without affecting machine learning.

[0126] Standardize the data according to the vectorization and standardization defined by the design side.

[0127] Mark data to indicate whether historical data has been scaled in or out, and also mark specific scaling indicators, such as the number of VNFC virtual machines to be scaled in or out.

[0128] Split the historical data into training data and test data according to the designed split ratio, such as 3:2.

[0129] Step 7: Feature Engineering.

[0130] Feature engineering is divided into dimension splitting and data standardization.

[0131] Dimension splitting: split a feature into multiple dimensions. For example, if memory is a feature, it can be split into five dimensions: within 512M, 512M-1G, 1G-2G, 2G-4G, and 4G-8G.

[0132] Data normalization is to normalize data of all dimensions according to the designed method of 0 / 1 or -1 / +1.

[0133] For example, if the current memory consumption is 800M, then only the dimension value of 512M-1G is 1, and the others are all 0 (if the design is +-1, then it is -1).

[0134] Step 8: Machine Learning.

[0135] When the data is ready, machine learning is performed according to the model, optimization algorithm, and hyperparameters defined in the design phase. Classification model learning is performed first, followed by regression model learning, to obtain the optimal solution and evaluation results.

[0136] If the evaluation effect is not good enough, then you need to readjust the parameters, that is, expand the amount of training data, redesign the features, or reselect the algorithm and hyperparameters.

[0137] Once the evaluation is greater than the minimum design requirement, the learning is completed and the model is output.

[0138] Please note that the parameter adjustment process requires human participation.

[0139] Step 9: Release the machine learning model.

[0140] To publish the learned model, you can save it as a file in a standard format and place it in a specified location.

[0141] Step 10: Convert the machine learning model into a slicing and scaling model through the model converter.

[0142] By design definition,

[0143] 1. Whether it is a fault - Determine whether it is a fault based on the 5G slice service quality KQI and 5G slice network performance KPI indicators. If so, hand it over to the fault ticket system for processing (the fault ticket system is not within the scope of this invention);

[0144] 2. Determine whether scaling is necessary based on the 5G slicing service experience KQI indicator

[0145] 3. If scaling is required, the converter calculates the adaptive scaling index of the eMBB slice based on real-time data and the regression model, and outputs the corresponding strategy according to the index and design definition, such as expanding the UPF instance.

[0146] Step 11: Strategy execution.

[0147] During real-time operation, the system will scale up and down according to the strategy, increase UPF instances, and directly increase the corresponding VNFC virtual machines based on resource data.

[0148] Figure 2 The entire system architecture and process of the method for implementing network-aware slice elastic scaling of the present invention are shown, which is divided into a design domain and an execution domain.

[0149] The design domain is designed and previewed on AIE Studio.

[0150] It needs to design the data, features, machine learning operators, and effect evaluation required for machine learning. It also needs to design how to convert the machine learning model into a strategy model and define the entire process.

[0151] Defining the process involves designing an entire semi-automated process. Parameters need to be readjusted only when the machine learning effect evaluation fails to meet the standards. All other processes are automated. That is, after data preparation, it is output to feature engineering to define features. After the features are defined, classification and regression model learning are performed. The learning results are directly evaluated. If they meet the standards, the machine learning model is automatically released. The network perception degradation index and adaptive scaling index are then calculated using the machine learning model and converted into a policy model through a converter.

[0152] The execution domain is to semi-automatically execute the entire process according to the designed process.

[0153] Figure 3 The dynamic scaling process of the present invention is shown, including: when a runtime data is collected, it is cleaned in the data preparation stage, and then feature engineering is performed to obtain feature data. It is judged whether there is a fault based on the 5G slice service quality KQI and 5G slice network performance KPI indicators. If so, it is handed over to the fault work ticket system for processing (the fault work ticket system is not within the scope of the present invention); if it is not a fault, it is judged whether scaling is required based on the 5G slice service experience KQI indicator combined with the knowledge base in the knowledge graph. Note that the expert knowledge base can be a calculation formula for each KQI indicator or an empirical rule; if scaling is required, it is calculated using the operator of the machine learning regression model, and the output is passed through the model converter to obtain how the current strategy is scaled, and then the scaling is specifically executed.

[0154] For example, a popular VR anchor will broadcast at 20:00 on May 1st.

[0155] 1) The current 5G slice service quality KQI video download rate and video rate bit rate ratio have both reached the upper limit of 80%, and the 5G slice network performance KPI data throughput has reached the upper limit of 90%, indicating that there is no equipment failure. If there is a failure, the network data or download data will be abnormal;

[0156] 2) Based on the 5G end-side service experience KQI indicators, if the duration of the interruption indicator exceeds the average interruption time and the packet loss rate exceeds 20%, the knowledge graph determines whether scaling is required;

[0157] 3) Based on the feature data and the regression model, the adaptive shrinkage index value 8 of the eMBB slice is obtained.

[0158] 4) Converted into a policy model, the UPF is scaled up or down by 8 times its original initial value, that is, the corresponding VNFC is orchestrated and executed to scale up or down to 8 virtual machines.

[0159] Figure 5 The flowchart of the execution engine of the present invention is shown. The execution of the semi-automatic process of the present invention is driven by the execution engine.

[0160] The execution engine will execute according to the designed process, obtain the current process, and get the input data, convert the input data into the data required by this process (optional), and then execute the process and output it.

[0161] For example, the current process is classification model learning, and its input data is the data after feature engineering. It will perform machine learning based on the designed and defined learning operators and then output the results.

[0162] If the current process is an evaluation process, then its input data is the data after machine learning. It will be evaluated according to the designed evaluation definition. If the goal is not achieved, the execution engine will be terminated and the parameter design will be readjusted.

[0163] If the evaluation reaches the goal, the process continues to the next step. If the process ends, the final policy model is output.

[0164] Figure 6 This section shows the feature definition and feature engineering process of the present invention. Feature engineering is a necessary step before machine learning. On the design side, features are designed, determining which data can be used as feature data, whether this data is discrete or linear, and how to divide this data into several dimensions. For example, the number of online users is not easy to divide using discrete data, so it can be treated as continuous data. The distribution and variance are then statistically analyzed, and the dimensions are divided according to the different variances.

[0165] Although the present invention has been described in detail above, it is not limited thereto, and those skilled in the art can make various modifications based on the principles of the present invention. Therefore, any modifications made based on the principles of the present invention should be understood to fall within the scope of protection of the present invention.

Claims

1. A network-aware slice elastic scaling method, characterized in that: include: Real-time collection of business experience KQI indicators; Determine whether the network is normal based on the collected service experience KQI indicators; When the network is judged to be normal, the system determines whether the slice needs to be scaled based on the collected service experience KQI indicators and the knowledge in the expert knowledge base. If the slice needs to be scaled, the system performs the scaling operation on the slice. The scaling operation on the slice includes: The number of slices to be scaled is determined using a slice scaling model and real-time network data. The slice scaling model is a linear regression model for slice scaling, which is obtained by converting historical data into feature data and then performing machine learning. The real-time network data includes KPI indicators such as traffic, CPU and memory, as well as real-time service data. Increase or decrease slices according to the number of slices that need to be scaled; The slice scaling model and real-time network data are used to determine the number of slices that need to be scaled, including: After converting the network real-time data into feature data, substituting the data into a linear regression model to obtain a first value; wherein the first value is calculated based on the weights of the feature data and the feature dimensions corresponding to the feature data; Dividing the first value by the initial value when not stretched to obtain a second value; Divide the second value by n and round it up to obtain the multiple required for scaling; wherein n is a positive integer representing the number of feature data; Multiply the required scaling factor by the current slice data to obtain the slice data that needs to be scaled.

2. The method according to claim 1, characterized in that The real-time collection service experience KQI indicator includes: real-time collection slice service quality KQI and slice network performance KPI.

3. A device for slice elastic scaling based on network awareness, characterized in that: include: Collection module, used to collect business experience KQI indicators in real time; The network fault judgment module is used to judge whether the network is normal based on the collected service experience KQI indicators; The slice scaling determination module is used to determine whether a slice needs to be scaled based on the collected service experience KQI indicators and the knowledge of the expert knowledge base when the network is judged to be normal, and to perform the scaling operation on the slice if it is determined that the slice needs to be scaled; The slice scaling determination module includes: The slice scaling model is used to determine the number of slices to scale based on real-time network data. This model is a linear regression model for slice scaling, derived by converting historical data into feature data and then performing machine learning. The real-time network data includes KPIs for traffic, CPU, and memory, as well as real-time service data. A scaling execution unit, configured to increase or decrease slices according to the determined number of slices that need to be scaled; wherein the real-time network data is converted into feature data and substituted into a linear regression model to obtain a first value; wherein the first value is calculated based on the weights of the feature data and the feature dimensions corresponding to the feature data; Dividing the first value by the initial value when not stretched to obtain a second value; Divide the second value by n and round it up to obtain the multiple required for scaling; wherein n is a positive integer representing the number of feature data; Multiply the required scaling factor by the current slice data to obtain the slice data that needs to be scaled.

4. The device according to claim 3, characterized in that The real-time collection service experience KQI indicator includes: real-time collection slice service quality KQI and slice network performance KPI.

Images