INT and Machine Learning Based Service Quality Prediction Method

Abstract

The present invention discloses a method for predicting quality of service based on INT and machine learning, comprising the steps of: S1, a user terminal initiates an Internet access request, and a controller instructs the user terminal to send a data packet according to the user request; S2, switches at each node in the network perform in-band telemetry according to the telemetry instruction field in the data packet, collect INT metadata and embed it into the header of the data packet; S3, the last hop switch separates the INT metadata from the user data packet, the terminal telemetry server extracts the INT metadata, and uses a machine learning training method to establish a prediction model; S4, performs a service quality evaluation based on the output prediction model; S5, based on the service quality evaluation result, the background management terminal issues instructions to the controller or server to adjust resource allocation. The present invention more accurately predicts the status of services running in the network while ensuring the integrity of the data packet by embedding metadata in the header of the data packet through in-band telemetry.

Description

Technical Field

[0001] The present invention relates to a network security prediction method, and in particular to a service quality prediction method based on INT and machine learning. Background Art

[0002] With the advent of the information age, network usage has surged, and the number of network services requested by users has increased dramatically in the past decade. However, the complex network environment of the Internet makes the status of services ever-changing, so effective service quality prediction is very important at this time. Traditional service quality prediction methods mainly extract the required information from the message. However, the current traditional methods are not perfect in terms of service quality prediction, and there are still the following problems:

[0003] 1) Traditional service prediction uses limited packet data, including only surface information such as source address, destination address, source port, destination port, and protocol type. It cannot conduct in-depth analysis of the internal data packets, which affects the accuracy of service quality prediction and thus affects the user experience of network services.

[0004] 2) Currently, there are some AI-based time series prediction models for service quality, such as the long short-term memory model (LSTM). These models combine neural networks with prediction algorithms to improve the prediction ability of network service quality. However, due to the limitations of neural networks themselves, these models still have certain prediction errors when facing complex network scenarios.

[0005] In-band network telemetry (INT) is a new network measurement method. INT can quickly collect and integrate network status data to monitor service quality, and can deeply read network status for analysis to achieve end-to-end visualization of network services. The machine learning method can extract features and perform cluster analysis on network data in the data preprocessing stage, and improve the existing prediction model to improve the prediction accuracy of service quality. Summary of the invention

[0006] Purpose of the invention: The purpose of the present invention is to provide a service quality prediction method based on INT and machine learning that can improve prediction speed and accuracy.

[0007] Technical solution: The service quality prediction method of the present invention comprises the following steps:

[0008] S1, the user terminal initiates an Internet access request, and the controller instructs the user terminal to send a data packet according to the user request;

[0009] S2, each node switch in the network performs in-band telemetry based on the telemetry instruction field in the data packet, collects INT metadata and embeds it into the header of the data packet;

[0010] S3, the last-hop switch separates the INT metadata from the user data packet, the terminal telemetry server extracts the INT metadata, uses machine learning methods to preprocess the data, and uses the ANFIS neural network to build a model;

[0011] S4, evaluate the service quality based on the output prediction model;

[0012] S5, based on the service quality evaluation result, the backend management terminal sends instructions to the controller or server to adjust resource allocation.

[0013] Furthermore, in step S3, after the metadata is input into the training layer, a prediction model is output through multiple iterations of training of the neural network training module, and the prediction model is stored.

[0014] Further, in step S3, the quality evaluation index result of the telemetry metadata INT measurement is extracted as data set X, and the QoS measurement of the user data packet is data set Y; based on the fine-grained telemetry metadata {X t}, estimate the network service quality metric {Y t}; The detailed implementation steps are as follows:

[0015] S31, divide the data set X into n non-overlapping feature spaces r1, r2, ...r i , the Qos metric Y is divided into n different feature spaces y1, y2, ...y i , 1≤i≤n, calculate the maximum likelihood distance between the two feature spaces, and extract the common features h(x,y) of the data sets in the two feature spaces:

[0016]

[0017] Cluster the extracted common features h(x,y) and input the dataset X for neural network training train , grouping the fine-grained attributes such as delay and traffic in the common feature h(x,y), data set X train The expression is as follows:

[0018]

[0019] Among them, μ is the membership function, n is the number of feature spaces, and m is the maximum number of groups;

[0020] S32, training data set X train Input the deep learning ANFIS neural network for k iterations of training, where the number of iterations is determined by the dataset X train The size of the output initial prediction function Then there is

[0021]

[0022] Among them, w l represents the weight of the neural network trained at the lth iteration, 1≤l≤k; σ is the activation function used to train the telemetry metadata X t Mapping of data features;

[0023] S33, for the initial prediction function Perform error correction to obtain the final prediction function F(X t );

[0024] S34, according to the final prediction function F(X t ), output prediction model

[0025] Further, in step S33, the quadratic error function is optimized until the error accuracy meets the output requirements to obtain the final prediction function F(X t ), then:

[0026]

[0027] Among them, λ is a constant, ε j 2 is the quadratic error function, obtained from the jth training And j-1 times training The difference.

[0028] Further, in step S4, the service quality evaluation is performed by recalculating the network service quality metric Y t The QoS metric Y of the user data packet test The maximum likelihood estimate μ is used to reflect the accuracy of the prediction model, and the expression is:

[0029]

[0030] Among them, m is the maximum number of groups.

[0031] Compared with the prior art, the present invention has the following significant effects:

[0032] 1. The present invention adopts an in-band telemetry method to collect data. By embedding metadata in the header of a data packet through in-band telemetry, the data can be collected in a more fine-grained manner while maintaining the integrity of the data packet. This can more deeply predict the service quality of the current application from the underlying data;

[0033] 2. Compared with traditional time - series - based prediction models, it has a faster prediction speed and higher accuracy. On the one hand, the large number of network metrics allowed by INT brings a new view of the network to network administrators, enabling fine - grained data collection, such as the path of data packets, the occupancy rate of buffers, and the waiting time of data packets in the queue. On the other hand, machine - learning algorithms obtain telemetry metadata and improve existing prediction models based on machine - learning feature extraction and clustering analysis in the pre - processing stage. By inputting the training data set, the prediction accuracy of the model is improved. The established prediction model itself can handle high - dimensional network features, extract hidden features, and has good anti - noise performance. It can more accurately predict the state of services running in the network while ensuring the integrity of data packets, thus providing opportunities for new management operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is the platform architecture diagram of the present invention;

[0035] Figure 2 It is the full - process flowchart of the prediction of the present invention;

[0036] Figure 3 It is the overall flowchart of neural network training. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The present invention will be further described in detail below in conjunction with the accompanying drawings of the specification and the specific embodiments. The described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present invention.

[0038] The present invention first uses in - band telemetry for data collection. After the telemetry server parses the telemetry metadata, it inputs it into the training layer, uses machine - learning - related algorithms for data pre - processing to improve the model training accuracy, then uses the ANFIS neural network for training, and finally evaluates based on the output prediction model. This method improves the accuracy of service - quality prediction and helps the server - side further process to improve the user experience. To achieve the above - mentioned purpose, as Figure 1 shown, the present invention includes the following parts: a user layer, a training layer, and a service layer.

[0039] The user layer represents the requester of the current network service. After the service requester initiates a service request, the controller or server issues an instruction to send a data packet. The in-band telemetry module is deployed in the user layer. The working principle of in-band telemetry is that the switch telemetry instruction embeds a telemetry header in the user data packet to form telemetry metadata. The data packet carries the telemetry data of each node switching device hop by hop and is forwarded by the last-hop switch to the telemetry server. The telemetry server extracts the telemetry metadata for analysis. The telemetry instruction is issued by the controller together when the service is requested. After the data packet arrives at the receiving end, the telemetry server at the receiving end separates the user data packet and the telemetry metadata header. The metadata parsing module in the user layer parses the telemetry metadata, obtains network states such as the delay and bandwidth of the current network, and forwards them to the training layer.

[0040] The function of the training layer is to collect telemetry metadata, obtain various indicators such as delay and bandwidth in the metadata, preprocess it in the data preprocessing module, and then input it into the neural network training module for training. The data preprocessing module mainly extracts and aggregates features of the collected network state parameters through machine learning algorithms to construct a training data set, so as to improve the quality of the input data and enhance the training effect of the neural network module. The neural network training module inputs the data set of the preprocessing module into the neural network and uses the neural network training module for multiple iterative trainings until the output data accuracy meets the corresponding requirements, saves the trained prediction model. The prediction model generated after the training data set preprocessed by machine learning algorithms is input into the neural network can itself process high-dimensional network features, extract hidden features, and has good anti-noise performance. It can more accurately predict the state of the service running in the network while ensuring the integrity of the data packet, thus providing opportunities for new management operations.

[0041] The main function of the service layer is to evaluate the service quality according to the various indicators fed back by the prediction model and generate a visual interface of various service indicators on the management end monitoring interface. The service layer uploads the service quality evaluation results to the management end. According to the evaluation results of various indicators, the background management end issues instructions to the controller or server to adjust resource allocations such as bandwidth and traffic, so as to ensure the network service usage experience of the user side.

[0042] The present invention uses in-band telemetry to collect network data. On the premise of maintaining the integrity of the data packet, the telemetry server parses the metadata, preprocesses it through machine learning algorithms to obtain a training data set for model training, inputs it into the neural network model for training, and finally conducts evaluation and prediction based on the training model.

[0043] As Figure 2 shown, the overall process of the service quality prediction of the present invention is as follows:

[0044] Step 1, the user side initiates an Internet access request, and the controller instructs the user side to send a data packet according to the user request;

[0045] Step 2: Each node switch in the network performs in-band telemetry based on the telemetry instruction field in the data packet, and collects INT metadata to embed it into the header of the data packet;

[0046] Step 3: The last-hop switch separates the INT metadata from the user data packet, and the terminal telemetry server extracts the INT metadata, uses machine learning methods for data preprocessing to improve the fine-grainedness of the prediction model, and establishes a model using an ANFIS neural network.

[0047] After the metadata is input into the training layer, it will go through multiple iterative trainings of the neural network training module to output a prediction model and store the model for evaluation and prediction in the service layer. As Figure 3 shown in the overall flowchart of neural network training.

[0048] Let the time delay result of extracting the telemetry metadata tag INT metric be the data set X, and the Qos metric of the user data packet be the data set Y. The Qos metric is a reflection of the network quality. Since the network state changes in real time, the prediction needs to be measured at each moment t. The purpose is to estimate the network service quality metric {Y t} based on the fine-grained telemetry metadata {X t} within the time t, and finally output a service quality prediction model of {X t} → {Y t}.

[0049] The detailed implementation steps are as follows:

[0050] Step 31: Data preprocessing

[0051] In the data preprocessing stage, the present invention will use the fuzzy clustering algorithm in machine learning to extract features and cluster the collected telemetry metadata, identify similar features among different telemetry metadata, and accumulate similar objects into the same group as the data set for model training.

[0052] Divide the data set X into n non-overlapping feature spaces r1, r2,..., r n , and denote the i-th feature space as r i (1 ≤ i ≤ n). Divide the Qos metric Y into n different feature spaces y1, y2,..., y n , and also denote the i-th feature space as y i (1 ≤ i ≤ n). Calculate the maximum likelihood distance between the two feature spaces, and extract the common feature h(x, y) of the data sets in the two feature spaces:

[0053]

[0054] Cluster the extracted common features h(x,y), and input the dataset X for neural network training train , as shown in formula (2), μ is the membership function, which can group fine-grained attributes such as time delay and traffic in the common feature h(x,y) and associate them with the label i in the feature space:

[0055]

[0056] Among them, n is the number of feature spaces, and m is the maximum number of groups (i.e., the maximum number of groups when performing grouping of fine-grained attributes).

[0057] Step32, Input the preprocessed training dataset into the ANFIS neural network for model training.

[0058] Input the training dataset X train into the deep learning ANFIS neural network for k iterations of training. The number of iterations is determined by the size of the dataset X train . The larger the dataset, the higher the number of iterative training times is required to ensure the accuracy of the model, and the initial prediction function is output Then there is

[0059]

[0060] Among them, w l represents the weight of the neural network in the l-th iteration of training; σ is the activation function, which is used to map the data features of the telemetry metadata X t .

[0061] Step33, Error correction

[0062] Since the neural network is sensitive to parameter changes, a quadratic error function needs to be added for optimization after training the training model based on formula (3) until the error accuracy meets the output requirements to obtain the final prediction function F(X t ), then there is:

[0063]

[0064] Among them, λ is a constant, and ε j 2 is the quadratic error function, which is the difference between the obtained in the j-th training and the obtained in the (j - 1)-th training .

[0065] Step34, Output the prediction model after training

[0066] According to the final prediction function F(X t ), output the prediction model

[0067] Step 4: Perform service quality evaluation based on the output prediction model, and input the test data set Y test Conduct an evaluation. The evaluation basis is obtained by recalculating the test data set Y test (i.e., the Qos metric of the user data packet) and the maximum likelihood estimate μ of the training data set Y t (i.e., the network service quality metric) to reflect the accuracy of the prediction model. It represents the distance between the predicted value and the actual value. Generate a visualization interface for the delay metrics included in the prediction model in the management - end monitoring interface. According to different input values, the prediction model can record the corresponding network status from different aspects (such as reliability metrics like delay, and robustness - reflecting metrics like throughput). As shown in formula (5):

[0068]

[0069] Step 5: According to the evaluation results, the background management end issues instructions to the controller or server to adjust resource allocations such as bandwidth and traffic, so as to ensure the network service usage experience of the user end.

Claims

1. A service quality prediction method based on INT and machine learning, characterized in that, The steps include: S1, the user terminal initiates an Internet access request, and the controller instructs the user terminal to send a data packet according to the user request; S2, each node switch in the network performs in-band telemetry based on the telemetry instruction field in the data packet, collects INT metadata and embeds it into the header of the data packet; S3, the last-hop switch separates the INT metadata from the user data packet, the terminal telemetry server extracts the INT metadata, uses machine learning methods to preprocess the data, and uses the ANFIS neural network to build a model; S4, evaluate the service quality based on the output prediction model; S5, based on the service quality evaluation results, the backend management terminal sends instructions to the controller or server to adjust resource allocation; In step S3, after the metadata is input into the training layer, the prediction model is output through multiple iterations of training of the neural network training module, and the prediction model is stored; The quality evaluation index results of the telemetry metadata INT measurement are extracted as dataset X, and the Qos measurement of the user data packet is dataset Y; Estimate the network quality of service metric {Y t} based on the fine-grained telemetry metadata {X t} within time t; the detailed implementation steps are as follows: S31. Divide the data set X into n non - overlapping feature spaces r1, r2, … rn i , and divide the Qos metric Y into n different feature spaces y1, y2, … yn i , where 1 ≤ i ≤ n, calculate the maximum - likelihood distance between the two feature spaces, and extract the common feature h(x, y) of the data sets of the two feature spaces: Cluster the extracted common feature h(x,y), and input it into the dataset X for neural network training train , group the delay and traffic in the common feature h(x,y), and the dataset X train is expressed as follows: Among them, μ is the membership function, n is the number of feature spaces, and m is the maximum number of groups; S32, input the training data set X train into the deep learning ANFIS neural network for k - iteration training, where the number of iterations is determined by the size of the data set X train and output the initial prediction function Then there is Among them, w l represents the weight of the neural network in the l-th iterative training, where 1 ≤ l ≤ k; σ is the activation function used to map the data features of the telemetry metadata X t ; S33. Perform error correction on the initial prediction function to obtain the final prediction function F(X t ); S34, according to the final prediction function F(X t ), output the prediction model 2. The service quality prediction method based on INT and machine learning according to claim 1, wherein In step S33, quadratic error function optimization is adopted until the error precision meets the output requirements to obtain the final prediction function F(X t ), then we have: where λ is a constant, and ε j 2 is the quadratic error function, and is the one obtained from the j-th training minus the one obtained from the (j - 1)-th training difference.

3. The quality of service prediction method based on INT and machine learning according to claim 1, characterized in that, In step S4, the quality of service evaluation is performed by recalculating the network quality of service metric Y t and the QoS metric Y of the user data packet test of the maximum likelihood estimate value μ to reflect the accuracy of the prediction model, and the expression is: Among them, m is the maximum number of groups.

Images