CN111526070A - Service function chain fault detection method based on prediction - Google Patents

Abstract

The invention relates to a prediction-based service function chain fault detection method, which belongs to the technical field of communication. The present invention firstly collects data by monitoring the performance data of each virtual network function on the entire service function chain according to the performance correlation existing between the virtual network functions of the service function chain under the virtual network environment, and divides its working state; Secondly, considering the high-dimensional complexity and time-related characteristics of network monitoring data, combined with the initiative requirements of fault detection, a gated recurrent unit network is used to detect faults, and the health status of the network is predicted by analyzing the historical performance data information of the service function chain. ; And using the similarity of virtual network function nodes between different service function chains, transfer learning is introduced to speed up the convergence of the model. The present invention can effectively detect the occurrence of a fault while meeting the fault detection time requirement.

Description

A prediction-based service function chain fault detection method

technical field

The invention belongs to the technical field of communication, and relates to a prediction-based service function chain fault detection method.

Background technique

The arrival of 5G will bring about a change across the ages. 5G networks will provide differentiated business scenarios and customized network services for people's diverse service needs. The existing network architecture is mainly oriented to the communication between people and people, and it is difficult to support emerging services such as the Internet of Things and the Internet of Vehicles that require large-scale machine communication. Therefore, 5G regards network slicing as an ideal network architecture as the key to solving the above problems. The 5G network slicing architecture based on SDN and NFV can be flexibly arranged according to user needs and allocate resources dynamically and efficiently. NFV technology decouples the software and hardware of traditional networks and implements network functions in software, which is more conducive to the sharing of network resources.

However, while the network slicing architecture brings great flexibility to the 5G network, it also puts forward new requirements for the reliability of the network. Due to the sharing of the underlying resources, the failure of the underlying network node can easily lead to the failure of multiple Virtual Network Functions (VNFs) sharing the resources of the node, resulting in functional paralysis of multiple Service Function Chains (SFCs). Therefore, it is very important to use the SON technology to realize the rapid recovery of the network from the fault, and the fault detection, as the main body of the network performance analysis, is the primary prerequisite for realizing self-healing measures.

In the process of researching the prior art, the inventor found that it has the following shortcomings:

Most of the existing fault detection methods are based on the reaction mechanism. However, this mechanism is difficult to avoid the inherent delay of the reaction, which is not conducive to the timeliness requirements of the network; most methods do not consider the degradation of network service quality, and ignore the continuous performance degradation. Adverse effects on the network; failures in the scenario of service function chains without considering the vertical propagation between the infrastructure layer and the application layer and horizontal propagation between different service function chains will cause some normal VNF nodes to appear. In the case of abnormal performance, it is difficult to find the location of the VNF node where the fault first occurred.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide a prediction-based service function chain fault detection method, which can effectively detect node anomalies according to changes in the performance data of VNF nodes in a network virtualization environment, so as to meet the network requirements. reliability requirements.

To achieve the above object, the present invention provides the following technical solutions:

A prediction-based service function chain fault detection method, characterized in that the method specifically includes the following steps:

S1: Combined with the characteristics of fault propagation in the service function chain scenario, according to the performance correlation between VNF nodes, the data is collected by monitoring the performance data of each VNF on the entire service function chain, and its working status is divided into normal, service quality degradation or malfunction;

S2: In view of the high-dimensional complexity and time-related characteristics of network monitoring data, combined with the initiative requirements of fault detection, a gated recurrent unit (GRU) network is used to detect faults, and the historical performance data of the service function chain is analyzed by analyzing the historical performance data. information to predict the health of the network; among them, for the problem that GRU network modeling takes a long time and is not conducive to the real-time requirements of fault detection, the similarity of VNF nodes between different service function chains is used to introduce transfer learning to speed up the model convergence speed.

Further, in the step S1, in the service function chain scenario, the service function chain is usually formed by a plurality of VNFs with independent functions arranged in a certain order according to the specific needs of the user, and the VNFs in different service function chains may have Partially overlapping, this feature of the service function chain can easily cause a VNF fault to propagate to the associated VNF, which in turn leads to large-scale faults in the slicing network.

Considering that faults are propagated between different VNF nodes in the service function chain, the performance data is collected by monitoring the working status of each VNF node in the entire service function chain at the application layer of the slicing network. Continuity analysis detects the occurrence of failures and considers the first VNF node that exhibits failures as the origin of failures.

Further, in order to eliminate the influence of different dimensions in the original data samples, different data value ranges, and insignificant data change trends on model training, and at the same time to improve the accuracy of the model and the speed of network training, it is necessary to use linear maximum and minimum values. The method performs normalization preprocessing on the performance data collected from all VNF nodes, and its transformation function is:

Further, in the step S1, the working state of the service function chain is divided into three categories according to the cause of the failure:

(1) Normal: The network status is running well;

(2) Deterioration of service quality: the network load increases, traffic decreases, delay increases or packet loss occurs, but the VNF can still work; the specific reasons may be sudden changes in the surrounding environment, software and hardware problems, or insufficient resources. Return to normal working state in a short time or continue to deteriorate into failure;

(3) Failure: The network function is completely unusable, the delay becomes infinite, and corresponding services cannot be provided for users; in order to ensure the normal operation of the SFC where the VNF node is located, operations such as node migration or software and hardware device restarts need to be performed.

Further, the rules for executing node migration are: the faulty VNF node may be generated by a mutation of a normal VNF node, or the performance of a VNF node whose service quality is degrading may deteriorate to a certain degree or abruptly generated during the performance degradation. For a VNF node whose service quality is degrading, if it can adapt to the optimal adjustment of the system, it will return to a normal working state or deteriorate into a fault. Necessary healing measures are needed to achieve network function recovery.

Further, in the step S2, in view of the high-dimensional complexity and time-related characteristics of the network monitoring data, combined with the initiative requirement of fault detection, the GRU network is used to detect the fault, which specifically includes the following steps:

S201: Use the three-layer GRU unit to process the input time series sample data, and use the historical monitoring data set to train the model in small batches;

S202: After passing through the three-layer GRU network, a fully connected layer is used to integrate the feature information of the previous network to improve the learning ability of the network;

S203: Use the output of the fully connected layer as the input of the softmax classifier, and perform reverse supervised fine-tuning combined with the label data;

S204: Use real-time monitoring data to further optimize parameters.

Further, in the step S2, the health status of the network is predicted by analyzing the historical performance data information of the service function chain, which specifically includes:

Assume that the historical performance data is the waiting delay and processing delay; in the training phase, firstly collect the characteristics of the waiting delay and processing delay of all VNF nodes on the service function chain, and set a service function chain to be composed of m VNF nodes , record the monitoring data of all VNF nodes at each moment, define the sliding window length as d, then in the time range from td to t, the input data set of the network model is expressed as x={x _t , x _t-1 ,..., x _t-d+1 }, the dataset of all VNF nodes at time t is:

和

分别表示第m个VNF的等待时延和处理时延；in,

and

respectively represent the waiting delay and processing delay of the mth VNF;

Prediction method: The input of the GRU network model is time series data, so it is necessary to construct time series samples through the sliding window method, with the length d as the size of the sliding window, and move on the data set according to the time step h to obtain the sample X at the current moment. _t ={x _t ,x _t-1 ,...,x _t-d+1 } and the next time sample X _t+h ={x _t+h ,x _t+h-1 ,...,x _{t+h -d+1} }, and determine the label values x _t+1 and x _t+h+1 according to the actual state of the network at the next moment;

Then, according to the dimension of the GRU input, the training set and the test set are divided according to a certain proportion, and the model is trained in small batches. After passing through the three-layer GRU network, the feature information of the previous network is processed through a fully connected layer. Integration, improve the learning ability of the network, and finally use the output of the fully connected layer as the input of the softmax classifier to obtain the final prediction result; in order to prevent overfitting when training the network, Dropout regularization is used to discard the data generated during the training process. partially duplicated information;

The training of the network model is the process of continuously optimizing the model parameters. In order to reduce the error between the output result and the real network, the network parameters are iteratively updated using the back-propagation algorithm; the network weights are updated layer by layer by means of gradient descent. optimization, so that the value of the objective loss function is minimized; because the Adam algorithm has advantages in computational efficiency and convergence speed compared with other parameter optimization algorithms, the Adam algorithm with adaptive learning rate is used to speed up the convergence speed of the algorithm.

Further, the Adam algorithm dynamically adjusts the learning rate of each parameter by using the distance estimation of the gradient, which is suitable for large data sets and high-dimensional spaces. The update formula is:

和

分别为梯度的一阶距估计的偏置矫正和二阶距估计的偏置矫正，δ是一个平滑项。where θ is the iteration parameter, ε is the learning rate,

and

are the bias correction for the first-order distance estimation of the gradient and the bias correction for the second-order distance estimation, respectively, and δ is a smoothing term.

Further, in the step S2, transfer learning is introduced to speed up the convergence speed of the model, which specifically includes: in the network slicing scenario, the size of the monitoring data set is subject to the length of the slicing running time. For the service function chain in the early stage of slicing operation For example, there may be cases where the fault detection accuracy is not high due to insufficient monitoring data. It is necessary to introduce a parameter-based transfer learning method to speed up the convergence of the model, so that the fault detection model based on GRU neural network prediction can be used in smaller Under the condition of data samples, it maintains a high detection accuracy.

The selected source domain model parameters are trained from the service function chain data with similar performance index requirements to the target domain; then the fault detection model parameters in other service function chains with similar structure to the current service function chain are transferred to the current service function chain, Help the current service function chain fault detection model to achieve better training results; the specific steps are: use the historical data sample set of the service function chain SFC b to train the fault detection model of the GRU network of SFC b, and obtain the optimal parameter matrix for model convergence , take _Wi ^b as an example; let the migration ratio φ(t)∈(0,1) represent the migration degree of parameters from the SFC b model to the SFC a model, where φ(t)=1/t, increasing with time t It can be seen that the parameter matrix of SFC a is W _i ^a =φ(t)W _i ^b +(1-φ(t))W _i ^a , and the parameter at the initial moment is W _i ^a =W _i ^b , using SFC The data sample set of a is used for model training and fine-tuning, and the optimal _GRU network model parameters of SFC ^a are obtained.

The beneficial effects of the present invention are: aiming at the problem of fault detection in the 5G end-to-end network slicing scenario, the present invention can effectively extract the massive and high-dimensional data features in complex networks on the basis of satisfying the system's requirements for detection accuracy. At the same time, the timeliness of fault detection is ensured, and it has high application value in wireless communication systems.

Other advantages, objects, and features of the present invention will be set forth in the description that follows, and will be apparent to those skilled in the art based on a study of the following, to the extent that is taught in the practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the following description.

Description of drawings

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be preferably described in detail below with reference to the accompanying drawings, wherein:

1 is a schematic diagram of a scene where the present invention can be applied;

FIG. 2 is a state transition diagram of a node in an embodiment of the present invention;

3 is a fault detection model based on a GRU network in an embodiment of the present invention;

FIG. 4 is a schematic flowchart of a service function chain fault detection method according to an embodiment of the present invention.

Detailed ways

The embodiments of the present invention are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments are only used to illustrate the basic idea of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

Please refer to FIG. 1 to FIG. 4 . FIG. 1 is a schematic diagram of a scenario where the present invention can be applied. Referring to Figure 1, the application layer is mainly responsible for providing an ordered set of VNFs for each slice request to process services, and the infrastructure layer provides physical nodes of many types of resources such as computing resources, bandwidth resources, and storage resources that support the functional requirements of various slice networks. and link. The virtualization layer implements functions such as slice lifecycle management and network performance data monitoring through NFV MANO and SDN controller. The system generates a specific service function chain according to different business requests, so as to meet the service needs of users. The traffic, bandwidth, and delay requirements of different VNFs in each service function chain are also different, but there is a certain correlation between the resources required by two adjacent VNFs and the virtual links connecting them. In order to ensure network stability and service quality, it is necessary to monitor the node status of different VNFs in the entire service function chain to detect the occurrence of faults in time.

FIG. 2 is a state transition diagram of a node in this embodiment. The faulty VNF node may be caused by the mutation of the normal VNF node, or the performance of the VNF node whose service quality is degraded may deteriorate to a certain extent or be abruptly generated during the performance degradation. For a VNF node whose service quality is degrading, if it can adapt to the optimal adjustment of the system, it may return to a normal working state.

FIG. 3 is a fault detection model based on a GRU network in an embodiment of the present invention. Referring to Fig. 3, the fault detection model based on GRU network prediction of the present invention is composed of three layers of GRU units, a fully connected layer and a softmax classifier. Define each input sample at time t as the health state feature information x _t of all VNF nodes on the same service function chain, and through the training of the GRU network model, the hidden layer state h _t of the model at this time can be obtained, and then obtain The predicted state y _t+1 of each VNF. Since the state of the VNF at the next moment is jointly affected by the observation data at this moment and the state of the hidden layer at the previous moment, the GRU network can use the historical observation data to predict the health state of the future VNF nodes.

FIG. 4 is a schematic flowchart of a service function chain fault detection method according to an embodiment of the present invention. Specific steps are as follows:

Step 401: Initialize system parameters, including model parameters of each layer of the GRU network of SFC b and SFC a, the learning rate and the number of iterations k=0;

Step 402: Input historical performance data of all VNF nodes of SFC b and real-time performance data of all VNF nodes of SFC a before time t;

Step 403: perform normalization preprocessing on the input data;

Step 404: Construct time series input data according to the sliding window length and time step;

Step 405: Use the historical performance data of all VNF nodes of SFC b to train the GRU network model, perform reverse fine-tuning of the model through the Adam algorithm of the adaptive learning rate, and update the corresponding parameters;

Step 406: Determine whether the convergence condition is satisfied. If the convergence condition is not met, set k=k+1, and continue to perform step 405; otherwise, extract the latest model parameters and perform step 407;

Step 407: Migrate the model parameters extracted from the GRU network model of SFC b to the GRU network model of SFC a, and use the real-time performance data before time t in SFC a to further train the model;

Step 408: Determine whether the maximum number of iterations K is reached. If not, let k=k+1, go to step 407, otherwise go to step 409;

Step 409: Output the working status of each VNF node in SFC a at time t+1.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should all be included in the scope of the claims of the present invention.

Claims (9)

1. A prediction-based service function chain fault detection method, characterized in that the method specifically comprises the following steps: S1: Combined with the characteristics of fault propagation in the service function chain scenario, according to the performance correlation between virtual network function (Virtual Network Function, VNF) nodes, the data is collected by monitoring the performance data of each VNF on the entire service function chain, and the Its working status is divided into normal, service quality degradation or failure; S2: In view of the high-dimensional complexity and time-related characteristics of network monitoring data, combined with the initiative requirements of fault detection, a gated recurrent unit (GRU) network is used to detect faults, and the historical performance data of the service function chain is analyzed by analyzing the historical performance data. information to predict the health of the network; among them, for the problem that GRU network modeling takes a long time and is not conducive to the real-time requirements of fault detection, the similarity of VNF nodes between different service function chains is used to introduce transfer learning to speed up the model convergence speed. 2. A prediction-based service function chain fault detection method according to claim 1, characterized in that, in the step S1, in the service function chain scenario, the service function chain is composed of a plurality of The VNFs with independent functions are arranged in a certain order; the performance data is collected by monitoring the working status of each VNF node of the entire service function chain at the application layer of the slicing network. The failure occurs, and the first VNF node that exhibits the failure is regarded as the starting point of the failure. 3 . The prediction-based service function chain fault detection method according to claim 2 , wherein the performance data collected from all VNF nodes is normalized and preprocessed by using a linear maximum-minimum value method. 4 . 4. A prediction-based service function chain fault detection method according to claim 1, wherein in the step S1, the working states of the service function chain are divided into three categories according to the cause of the failure: (1) Normal: The network status is running well; (2) Deterioration of service quality: the network load increases, the traffic decreases, the delay increases or the packet is lost, but the VNF can still work; (3) Failure: The network function is completely unavailable, the delay becomes infinite, and corresponding services cannot be provided for users; perform node migration or restart software and hardware devices. 5. A prediction-based service function chain fault detection method according to claim 4, wherein the rule for executing node migration is: for a VNF node whose service quality is deteriorating, if it can adapt to the system's Optimal adjustment will restore to the normal working state or deteriorate into a fault, and it is necessary to restore the network function through healing measures. 6. A prediction-based service function chain fault detection method according to claim 1, wherein in the step S2, for the high-dimensional complexity and time-related characteristics of network monitoring data, combined with the initiative of fault detection According to the requirements, the GRU network is used for fault detection, which specifically includes the following steps: S201: Use the three-layer GRU unit to process the input time series sample data, and use the historical monitoring data set to train the model in small batches; S202: After passing through the three-layer GRU network, a fully connected layer is used to integrate the feature information of the previous network to improve the learning ability of the network; S203: Use the output of the fully connected layer as the input of the softmax classifier, and perform reverse supervised fine-tuning combined with the label data; S204: Use real-time monitoring data to further optimize parameters. 7. A prediction-based service function chain fault detection method according to claim 6, characterized in that, in the step S2, the health status of the network is predicted by analyzing historical performance data information of the service function chain, specifically comprising: : Assume that the historical performance data is the waiting delay and processing delay; in the training phase, firstly collect the characteristics of the waiting delay and processing delay of all VNF nodes on the service function chain, and set a service function chain to be composed of m VNF nodes , record the monitoring data of all VNF nodes at each moment, define the sliding window length as d, then in the time range from td to t, the input data set of the network model is expressed as x={x _t , x _t-1 ,..., x _t-d+1 }, the dataset of all VNF nodes at time t is:

in,

and

respectively represent the waiting delay and processing delay of the mth VNF; Prediction method: Construct time series samples through the sliding window method, take the length d as the size of the sliding window, move on the data set according to the time step h, and obtain the current sample X _t ={x _t ,x _t-1 ,… ,x _t-d+1 } and the next moment sample X _t+h ={x _t+h ,x _t+h-1 ,...,x _t+h-d+1 }, and according to the next moment's sample The actual state of the network determines that the label values are x _t+1 and x _t+h+1 ; Then, according to the dimension of the GRU input, the training set and the test set are divided according to a certain proportion, and the model is trained in small batches. After passing through the three-layer GRU network, the feature information of the previous network is processed through a fully connected layer. Integration, improve the learning ability of the network, and finally use the output of the fully connected layer as the input of the softmax classifier to obtain the final prediction result; in order to prevent overfitting when training the network, Dropout regularization is used to discard the data generated during the training process. partially duplicated information; Training of the network model: iteratively update the network parameters by using the back-propagation algorithm; optimize the network weight layer by layer by gradient descent to minimize the value of the objective loss function; use the Adam algorithm with adaptive learning rate to speed up the algorithm. convergence speed. 8. a kind of prediction-based service function chain fault detection method according to claim 7, is characterized in that, Adam algorithm utilizes the distance estimation of gradient to dynamically adjust the learning rate of each parameter, and its update formula is:

where θ is the iteration parameter, ε is the learning rate,

and

are the bias correction for the first-order distance estimation of the gradient and the bias correction for the second-order distance estimation, respectively, and δ is a smoothing term. 9 . The prediction-based service function chain fault detection method according to claim 1 , wherein in the step S2, transfer learning is introduced to speed up the convergence speed of the model, which specifically includes: the selected source domain model parameters. 10 . Train from service function chain data with similar performance index requirements to the target domain; then transfer the fault detection model parameters in other service function chains with similar structure to the current service function chain to the current service function chain to help the current service function chain fail The detection model achieves better training effect; the specific steps are: use the historical data sample set of Service Function Chain (SFC) b to train the fault detection model of the GRU network of SFC b, and obtain the optimal parameter matrix for model convergence. Wi ^b ; let the migration ratio φ(t)∈(0,1) represent the migration degree of parameters from the SFC b model to the SFCa model, where φ(t)=1/t; the parameter matrix of SFCa is W _i ^a ₌ φ(t)W _i ^b +(1-φ(t))W _i ^a , the parameters at the initial moment Wi ^a ₌ Wi ^b , use the _SFCa data sample set to train and fine-tune the model to obtain the optimal SFCa The _GRU network model parameters Wi ^a' .

Images