Method for combining service classification and traffic time sequence prediction in wireless network

Abstract

The invention discloses a method for combining service classification and traffic time sequence prediction in a wireless network. The method comprises the following steps of: enhancing traffic prediction through service classification under the condition that the service analogy of each traffic sequence is known; enhancing traffic prediction through service classification under the condition that the service category to which each traffic sequence belongs is unknown; and the service classification accuracy is enhanced by using the flow prediction under the condition that the service category to which each flow sequence belongs is known. According to the invention, the accuracy of service classification can be enhanced, and the error of service time sequence prediction can be reduced, so that the network structure and parameters can be better deployed.

Description

Technical Field

[0001] The present invention belongs to the field of network technology, and particularly relates to a method for joint service classification and traffic time series prediction in a wireless network. Background Art

[0002] Wireless networks have witnessed rapid development in recent years. On the wireless access side, wireless networks provide users with convenient and high-speed access means. Typical representative networks include, but are not limited to: cellular networks, Wireless Local Area Networks (WLANs), SparkLink, Bluetooth, etc.

[0003] Wireless networks support a wide range of service types, including daily data transmissions such as file sharing and web browsing, online games with high real-time requirements, video stream services with high bandwidth demands, voice communication services such as VoIP phones, mobile office enabling remote work, and Internet of Things applications connecting smart homes and industrial sensors, etc. With the continuous evolution of wireless networks, the network can handle various traffic more flexibly and efficiently, ensuring user experience and service quality.

[0004] Time series prediction refers to analyzing historical data to identify the patterns of data changes over time, thereby predicting future data trends. This method can help network operators and service providers predict the traffic demands of different services at different time periods, and then make corresponding network resource adjustments and optimizations. With the development of technology, time series analysis has become an indispensable tool in multiple industries such as finance, meteorology, and supply chain management. For wireless network services, the network needs to process data traffic from multiple devices, and the patterns of this traffic may change due to time, location, and user behavior. Through time series prediction, network service providers can predict which time periods will have more user connections and which areas will have more data transmission demands, and thus perform network optimizations in advance, such as adding hotspots, adjusting channel widths, or optimizing signal coverage.

[0005] Machine learning is an extremely important branch in the field of artificial intelligence, which endows computer systems with the ability of self-learning and self-improvement. This ability is not achieved through traditional programming methods, but rather enables the computer to analyze a large amount of data using algorithms, learn, and identify patterns, trends, and regularities from it. With the recognition of patterns, the computer can make predictions or decisions on new, unseen data, and this ability has demonstrated great potential and value in multiple fields such as image recognition, speech recognition, natural language processing, medical diagnosis, and stock market analysis.

[0006] Machine learning techniques can be classified into several main types, including supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Each type is suitable for different types of problems and datasets. Supervised learning is applicable to scenarios with clear input-output relationships and labeled data, unsupervised learning is applicable to scenarios for exploring the intrinsic structure of data, semi-supervised learning is applicable to scenarios with scarce labeled data, and reinforcement learning is applicable to scenarios where learning strategies through interaction with the environment are required.

[0007] Currently, more and more wireless network applications and services have relatively stringent requirements for latency characteristics and transmission reliability, such as online games, virtual reality, medical care, industrial sites, etc. For this reason, ensuring latency performance and transmission reliability performance have become key technical goals for the next-generation wireless networks and have received extensive attention in the industry.

[0008] Technical principle of the prior art:

[0009] Time window method for time series prediction.

[0010] The time window method is a technique widely used in time series analysis. It captures the temporal characteristics in the data by dividing continuous time data into a series of fixed or sliding time periods. The core of this method lies in selecting an appropriate time range, called a window, which serves as the basis for data analysis.

[0011] As Figure 1 shown, select subsequences from the time series formed by historical data, use the previous part of the sequence as the input of the training set (prediction feature), and the latter part as the output of the training set (prediction target) and import them into machine learning. When performing the test of time series prediction, find a time series as the prediction feature at the time division point (current time point) according to the same dimension and import it into the learning machine as the test set.

[0012] Therefore, to form a prediction training set / test set, only the start time, the input length M of the training set / test set, the output length N of the training set / test set, and the sampling time step need to be selected. Then form a sequence with a length of M + N, use the first M sequence points as the input and the last N sequence points as the output. Repeat the above operations to form multiple training sets / test sets and merge them into the training set input matrix X i , output matrix Y i , test set input matrix X o , output matrix Y o .

[0013] Back Propagation (BP) neural network.

[0014] As Figure 2 shown in the BP neural network diagram, use represents the weight from the m-th input node of network layer l to the j-th hidden layer node of network layer l; use represents the weight from the j-th hidden layer node of network layer l to the p-th output node of network layer l; represents the offset of the j-th hidden layer node of network layer l; represents the offset of the p-th output node of network layer l; represents the output result of the p-th output node of layer l; is the input of the i-th input node of layer l, represents the output of the j-th hidden layer node of network layer l.

[0015] Considering the parameter update from the hidden layer to the output layer, then the output of the j-th hidden layer node of network layer l is:

[0016]

[0017] The output of the p-th neuron of network layer l is:

[0018]

[0019] where δ is the activation function, usually the sigmoid function:

[0020]

[0021] Its derivative is:

[0022] δ ′ (x) = δ(x)(1 - δ(x))

[0023] If the expected output is d = (d 1 , d 2 , … d p ) T , then the loss function is selected as:

[0024]

[0025] The goal of the BP neural network working is to minimize the loss function. If the difference between the actual output and the expected output is greater than the set threshold, then the value matrix is updated through a certain learning algorithm. In the standard BP neural network, the negative gradient is used for this process. The key to the gradient descent algorithm is to find the reciprocal of E p for each weight and bias term According to the chain rule of differentiation, we can get:

[0026]

[0027] Among them is the result of the transformation of the j-th hidden layer node in the l-th layer through ω and b, that is:

[0028]

[0029] Therefore, according to the above formula, the rule for updating the neural network parameters is:

[0030]

[0031] where σ is the learning rate set by net.trainParam.lr. The hidden layer error is:

[0032]

[0033] The rule for updating the parameters from the input layer to the hidden layer is the same as that from the hidden layer to the output layer.

[0034] When using machine learning for business time series prediction, the training set needs to be input into the learner to let the machine learn the patterns of the time series, and then the test set is used for prediction or model evaluation. Therefore, the effect of time series prediction depends on the quality of the training set. However, due to the increasing business complexity, the following problems exist in existing research:

[0035] 1) Most of the existing research on wireless network traffic prediction considers the total future traffic rather than distinguishing specific arriving services.

[0036] 2) Different services have different service models and arrival patterns. Therefore, when there are multiple service models in the training set, interference will occur at different times, reducing the prediction effect of machine learning. Summary of the Invention

[0037] In order to overcome the deficiencies of the prior art, the present invention provides a method for joint service classification and traffic time series prediction in a wireless network, which enhances traffic prediction through service classification when the service categories of each traffic sequence are known; enhances traffic prediction through service classification when the service categories to which each traffic sequence belongs are unknown; and enhances the accuracy of service classification by using traffic prediction when the service categories to which each traffic sequence belongs are known. The present invention can enhance the accuracy of service classification and reduce the error of service time series prediction, so as to deploy the network structure and parameters more optimally.

[0038] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0039] Step 1: Enhance traffic prediction through service classification when the service categories of each traffic sequence are known;

[0040] Step 1-1: Count all possible arriving services, and assign a numerical label to each service, denoted as l 1 and l 2 ...l n , where the numerical label can be a one-dimensional number or a multi-dimensional vector, and n is the total number of possible arriving services;

[0041] Step 1-2: Build predictors for all possible arriving services respectively;

[0042] Collect data for each possible arriving service to obtain a training set and a test set, and input them into a learning machine for learning to form net 1 and net 2 ...net n ; The features required to build the predictor include the following information:

[0043] · The time step unit for counting the total arriving traffic, that is, each sequence point of the time series is composed of the total arriving traffic in how long, which is called the sampling time step;

[0044] · The length of the input of the prediction training set / test set and the length of the output of the training set / test set;

[0045] · The true start time to the end time corresponding to each training set input sequence, which is called the input sequence corresponding time;

[0046] The input sequence corresponding time is equal to the product of the sampling time step and the length of the training set input;

[0047] Step 1-3: Build a classifier for all possible arriving services;

[0048] The features required to build the classifier include:

[0049] · The number of packets arriving within the input sequence corresponding time;

[0050] · The packet sending direction switching frequency within the input sequence corresponding time;

[0051] · The protocol type with the largest number of arrivals within the input sequence corresponding time;

[0052] · The protocol type with the second largest number of arrivals within the input sequence corresponding time;

[0053] · The full packet length within the input sequence corresponding time, that is, the packet length with the most occurrences;

[0054] · The average packet length within the input sequence corresponding time;

[0055] For the data within the input sequence corresponding time of a period, its classification label and the input of the prediction training set can be obtained at the same time, that is, its classification label can correspond to the predicted value;

[0056] Steps 1 - 4: When an actual business arrives, classify the characteristics of the business according to Steps 1 - 3 and input them into the classifier; output the business classification result x, and select the category label l 1 , l 2 ...l n The one closest to x among them is used as the classification result; The implementation methods for selecting the closest label include:

[0057] · Calculate the Euclidean distance between x and l 1 , l 2 ...l n and select the label with the minimum distance;

[0058] · Calculate the Hamming distance between x and l 1 , l 2 ...l n and select the label with the minimum distance;

[0059] · Calculate the Manhattan distance between x and l 1 , l 2 ...l n and select the label with the minimum distance;

[0060] · Calculate the cosine similarity between x and l 1 , l 2 ...l n and select the label with the maximum similarity;

[0061] Steps 1 - 5: Select the predictor corresponding to the business according to the business classification result, and divide the input of the test set according to the sampling time step, the time corresponding to the input sequence, and the time window method and input it into the predictor corresponding to the business for prediction;

[0062] Step 2: Enhance traffic prediction through business classification when the business category to which each traffic sequence belongs is unknown.

[0063] Step 2 - 1: Statistically analyze all the obtained traffic data to determine the input and output parameters of the classifier and the predictor, including the following aspects:

[0064] · Sampling time step;

[0065] · The length of the input of the prediction training set / test set and the output length of the training set / test set;

[0066] · The time corresponding to the input sequence.

[0067] Step 2-2: Extract the classification training set based on the obtained data. The method is to sample the existing business data, randomly select the start time to calculate the corresponding business classification features within the time period corresponding to the future input sequence, and the selection of business classification features is the same as in Step 1-3;

[0068] Step 2-3: Construct a business classifier, and select a clustering method to classify the sampling results of the classification training set in Step 2-2. The classification algorithms include:

[0069] · K-Means clustering method;

[0070] · Hierarchical clustering method;

[0071] · Fuzzy C-means calculation method;

[0072] Step 2-4: Classify the arriving business. When the business arrives, count the classification features of the business according to Step 1-3 and input them into the classifier to obtain the current business category;

[0073] Step 2-5: Extract the prediction training set based on the obtained data; the business data within the time period corresponding to the input sequence can be used simultaneously to generate classification labels, that is, the input of the classification training set and the prediction training set. The implementation method is as follows:

[0074] · Screen out the part of the business classification results in the sampling data of Step 2-2 that are the same as the current business category, and construct it according to the method of Step 1-2 based on this part of the business data;

[0075] · Sample the existing business data, randomly select the start time to calculate the corresponding business classification features within the time period corresponding to the future input sequence and input them into the business classifier, collect the business data that is the same as the current business category and construct it according to the method of Step 1-2;

[0076] Step 2-6: Use the generated prediction training set to construct a predictor, divide the input of the test set for the currently arriving business according to the time window method, and use the constructed predictor for prediction;

[0077] Step 3: Use traffic prediction to enhance the business classification accuracy when the business categories to which each traffic sequence belongs are known;

[0078] Step 3-1: Count all possible arriving businesses, and assign each business a digital label, denoted as l 1 、l 2 ...l n , where the digital label can be a one-dimensional number or a multi-dimensional vector, and n is the total number of possible arriving businesses;

[0079] Step 3-2: Construct predictors for all possible arriving businesses respectively;

[0080] Data collection is performed on each possible incoming service to obtain a training set and a test set, which are then input into a learner for learning to form net 1 、net 2 ...net n ; The features required to construct a predictor include the following information:

[0081] · The time step unit for counting the total arrival traffic, that is, each sequence point of the time series is composed of the total traffic that arrives in a certain period, which is called the sampling time step;

[0082] · The length of the input of the prediction training set / test set and the length of the output of the training set / test set;

[0083] · The true start time to the end time corresponding to each training set input sequence, which is called the time corresponding to the input sequence;

[0084] The time corresponding to the input sequence is equal to the product of the sampling time step and the length of the training set input;

[0085] Step 3-3: Construct a classifier for all possible incoming services;

[0086] The features required to construct a classifier include:

[0087] · The number of packets arriving within the time corresponding to the input sequence;

[0088] · The packet sending direction exchange frequency within the time corresponding to the input sequence;

[0089] · The protocol type with the largest number of arrivals within the time corresponding to the input sequence;

[0090] · The protocol type with the second largest number of arrivals within the time corresponding to the input sequence;

[0091] · The full packet length within the time corresponding to the input sequence, that is, the packet length with the most occurrences;

[0092] · The average packet length within the time corresponding to the input sequence;

[0093] For the data corresponding to the time of a segment of input sequence, its classification label and the input of the prediction training set can be obtained simultaneously, that is, its classification label can correspond to the predicted value;

[0094] Step 3-4: Sample the prediction test set according to the sampling time step, the time corresponding to the input sequence, and the time window method, and record its corresponding service category at the same time. Input it into each predictor for prediction and calculate the error E of each predictor according to the true value of the service sequence 1 、E 2 ...E n ;

[0095] Step 3-5: Import the currently arriving data into the classifier for classification, and calculate the current classification result \(x\) and each category label \(l\) 1 , \(l\) 2 ... \(l\) n The distance \(d\) 1 , \(d\) 2 ... \(d\) n , and the algorithm and steps for distance calculation are the same as those in Step 1-4;

[0096] Step 3-6: Set the weight factor \(\alpha\), where \(\alpha\) is a positive value; calculate the final result of the distance between the service category and each classifier label \([d\) 1 +\(\alpha E\) 1 , \(d\) 2 +\(\alpha E\) 2 ... \(d\) n +\(\alpha E\) n , select the service category corresponding to the minimum value as the final classification result, and compare it with the true service category;

[0097] Step 3-7: Repeat Step 3-4 to Step 3-6 to calculate the average classification accuracy rate, and improve the classification accuracy rate by adjusting the value of the weight factor \(\alpha\) and select the \(\alpha\) with the highest classification accuracy rate;

[0098] Step 3-8: After determining the weight factor \(\alpha\), input the true service sequence into each predictor and calculate the error \(E\) of each predictor respectively 1 , \(E\) 2 ... \(E\) n , and calculate the final classification result according to Step 3-5 and Step 3-6.

[0099] Preferably, the implementation method of the learner construction in the above Step 1-2 includes a statistical model, a machine learning model, a neural network model, and a hybrid model.

[0100] Preferably, in the above Step 3-4, the calculation method of the error includes:

[0101] · Root mean square error;

[0102] · Mean absolute error;

[0103] Preferably, in the above Step 3-7, the method for calculating \(\alpha\) is:

[0104] · Obtained according to the empirical value constructed by the service mathematical model;

[0105] · Traverse \(\alpha\) within a certain range and make a selection;

[0106] · Use machine learning to carry out parameter optimization work and select the best \(\alpha\).

[0107] A computer program that causes a computer to execute the above method for time series prediction.

[0108] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device executes the above method for time series prediction.

[0109] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the above method for time series prediction is implemented.

[0110] A chip includes: a processor, which is used to call and run a computer program from a memory, so that a device installed with the chip executes the above method for time series prediction.

[0111] A computer program product includes a computer storage medium, the computer storage medium stores a computer program, and the computer program includes instructions that can be executed by at least one processor. When the instructions are executed by the at least one processor, the above method for time series prediction is implemented.

[0112] The beneficial effects of the present invention are as follows:

[0113] The present invention can enhance the accuracy of service classification and reduce the error of service time series prediction, so as to deploy the network structure and parameters more optimally. Description of the Drawings

[0114] Figure 1 It is a schematic diagram of the time window method;

[0115] Figure 2 It is a structure diagram of a BP neural network;

[0116] Figure 3 It is a service sequence scenario with the problem of class mixing;

[0117] Figure 4 It is a diagram of Example 1;

[0118] Figure 5 It is a diagram of Example 2;

[0119] Figure 6 It is a diagram of Example 3. Detailed Embodiments

[0120] The present invention will be further described below with reference to the drawings and embodiments.

[0121] Due to the increasing complexity of wireless network services, there is a problem of class mixing in the network service traffic sequence: such as Figure 3As shown, network traffic consists of multiple types of services, such as video services, voice services, etc. Different services have different patterns. If professional means are not used to refine the screening and filtering of different categories during service collection, the effect of machine learning will be greatly reduced due to interference between different service patterns. For example, in the face of a voice service prediction task, the video service part in the training dataset will interfere with the machine learning effect.

[0122] The solution of the present invention is applicable to the prediction task of time series of service traffic with mixed service categories.

[0123] Embodiment 1: A general method for enhancing traffic prediction by service classification when the service categories of each traffic sequence are known;

[0124] Step 1: Count all possible arriving services and assign a digital label to each service, denoted as l 1 、l 2 ...l n , for example, let the digital label of WeChat voice service be 1, and the digital label of Huya live video service be 3. The label can be a one-dimensional number or a multi-dimensional vector. Where n is the total number of possible arriving services.

[0125] Step 2: Build predictors for all possible arriving services respectively. Collect data for each possible arriving service to obtain a training set and a test set, and input them into a learning machine for learning. Form net 1 、net 2 ...net n . The features required for building the predictor include the following information:

[0126] · The time step unit for counting the total number of arriving traffic, that is, each sequence point of the time series is composed of the total number of traffic arriving in how long, called the sampling time step. For example, if the sampling time step is set to 1s, the meaning of each sequence point in the service time series is the total service traffic (total packet length) arriving within 1s.

[0127] · The length of the input of the prediction training set / test set and the length of the output of the training set / test set.

[0128] · The true start time to the end time corresponding to each training set input sequence, called the input sequence corresponding time. The input sequence corresponding time is equal to the product of the sampling time step and the length of the training set input. For example, if the sampling time step is selected as 1s and the length of the training set input is 10, then the input sequence corresponding time is 1×10 = 10s.

[0129] The implementation method of building the learning machine can be composed of a statistical model, a machine learning model, a neural network model, a hybrid model, etc.

[0130] Step 3: Build classifiers for all possible arriving services. The features required for building classifiers may include but are not limited to the following information:

[0131] · The number of packets arriving within the time corresponding to the input sequence.

[0132] · The packet sending direction exchange frequency within the time corresponding to the input sequence.

[0133] · The protocol type with the largest number of arrivals within the time corresponding to the input sequence.

[0134] · The protocol type with the second largest number of arrivals within the time corresponding to the input sequence.

[0135] · The full packet length (the packet length with the most occurrences) within the time corresponding to the input sequence.

[0136] · The average packet length within the time corresponding to the input sequence.

[0137] Therefore, for the service data corresponding to a period of input sequence time, since the time corresponding to its input sequence is equal to the product of the sampling time step and the input length of the training set, the duration of the time corresponding to the input sequence considered for its classification label is the same as the total duration in the service traffic data corresponding to the input sequence considered for the prediction training set. Therefore, for the data corresponding to a period of input sequence time, its classification label and the input of the prediction training set can be obtained simultaneously, that is, its classification label can correspond to the predicted value. The implementation methods for building the learner can be composed of statistical models, machine learning models, neural network models, hybrid models, etc.

[0138] Step 4: When the actual service arrives, according to the classification features of the statistical service in Step 3, input them into the classifier. Output the service classification result x. Select the class label l 1 、l 2 ...l n The one closest to x among them is used as the classification result. The implementation methods for selecting the closest label include but are not limited to the following several:

[0139] · Calculate the Euclidean distance between x and l 1 、l 2 ...l n and select the label with the smallest distance.

[0140] · Calculate the Hamming distance between x and l 1 、l 2 ...l n and select the label with the smallest distance.

[0141] · Calculate the distance between x and l 1 、l 2 ...l nThe Manhattan distance and select the label with the smallest distance.

[0142] · Calculate the cosine similarity between x and l 1 、l 2 ...l n and select the label with the largest similarity.

[0143] Step 5: According to the business classification result, select the predictor corresponding to this business, and divide the input of the test set (prediction set) according to the sampling time step, the time corresponding to the input sequence, and the time window method, and input it into the predictor corresponding to this business for prediction.

[0144] Example 2: A general method for enhancing traffic prediction by business classification when the business category to which each traffic sequence belongs is unknown.

[0145] Step 1: Statistically analyze all the obtained traffic data, and determine the input and output parameters of the classifier and predictor, including the following aspects:

[0146] · Sampling time step.

[0147] · The length of the input of the prediction training set / test set and the output length of the training set / test set.

[0148] · The time corresponding to the input sequence.

[0149] Step 2: Extract the classification training set according to the obtained data. The method is to sample the existing business data, randomly select the start time, and calculate the corresponding business classification features within the time corresponding to the future input sequence. The selection of business classification features is the same as that in Step 3 of Example 1.

[0150] Step 3: Construct a business classifier. Since the specific business category is unknown, a clustering method is selected to classify the sampling results of the classification training set in Step 2. The classification algorithms include but are not limited to the following:

[0151] · K-Means clustering method.

[0152] · Hierarchical clustering method.

[0153] · Fuzzy C-means calculation method.

[0154] Step 4: Classify the arriving business. When a business arrives, count the classification features of the business according to Step 3 of Example 1 and input them into the classifier to obtain the current business category.

[0155] Step 5: Extract the prediction training set based on the obtained data. Since for the business data corresponding to the time of an input sequence, its classification label and the input of the prediction training set can be constructed simultaneously, the business data within the time corresponding to the input sequence can be used simultaneously to generate the classification label (classification training set) and the input of the prediction training set. Therefore, its implementation method can be as follows:

[0156] · Screen out the part of the business classification results in the sampled data of Step 2 that is the same as the current business category, and construct it according to the method in Step 2 of Embodiment 1 based on this part of the business data.

[0157] · Sample the existing business data, randomly select the start time to calculate the corresponding business classification features within the time corresponding to the future input sequence and input them into the business classifier, and collect the business data that is the same as the current business category and construct it according to the method in Step 2 of Embodiment 1.

[0158] Step 6: Use the generated prediction training set to construct a predictor, divide the input of the test set (prediction set) for the currently arriving business according to the time window method, and use the constructed predictor for prediction.

[0159] Embodiment 3: A general method for enhancing the business classification accuracy by using traffic prediction when the business category to which each traffic sequence belongs is known.

[0160] Steps 1 - 3 are the same as Steps 1 - 3 of Embodiment 1.

[0161] Step 4: Sample the prediction test set according to the sampling time step, the time corresponding to the input sequence, and the time window method (note that this test set does not really classify the actual business, but is used to select the weight factor α in Step 6), and record its corresponding business category at the same time. Input it into each predictor for prediction and calculate the error E of each predictor according to the true value of the business sequence 1 、E 2 ...E n . The calculation methods of the error include but are not limited to the following several types:

[0162] · Root mean square error.

[0163] · Mean absolute error.

[0164] Step 5: Import the currently arriving data into the classifier for classification, calculate the current classification result x and the distances d 1 、d 2 ...d n from the category labels l 1 、l 2 ...l n . The algorithm for distance calculation is the same as that in Step 4 of Embodiment 1.

[0165] Step 6: Set the weight factor α, where α should be a small positive value. Calculate the final result [d 1 + αE 1 , d 2 + αE 2 ... d n + αE n of the distance between the business category and the labels of each classifier, select the business category corresponding to the minimum value as the final classification result, and compare it with the true business category.

[0166] Step 7: Repeat Steps 4 to 6 to calculate the average classification accuracy rate, and improve the classification accuracy rate by adjusting the value of the weight factor α and select the α with the highest classification accuracy rate. The ways to calculate α can be:

[0167] · Obtained according to the empirical values constructed by the business mathematical model.

[0168] · Traverse α within a certain range and make a selection.

[0169] · Use machine learning to carry out parameter optimization work and select the best α.

[0170] Step 8: After determining the weight factor α, input the true business sequence into each predictor and calculate the error E 1 、E 2 ... E n respectively, and calculate the final classification result according to Steps 5 and 6.

Claims

1. A method for joint service classification and traffic time series prediction in a wireless network, characterized in that: The steps include: Step 1: Enhance traffic prediction by service classification when the service analogy of each traffic sequence is known; Step 1-1: Count all possible services that may arrive and give each service a numerical label, denoted as l1, l2, ...l n , the digital label can be a one-dimensional number or a multi-dimensional vector, where n is the total number of services that may arrive; Step 1-2: Build predictors for all possible arriving services; Data collection is performed on each possible business to obtain training sets and test sets, which are then input into the learner for learning to form net1, net2...net n ; The features required to build a predictor include the following information: The time step unit for counting the total number of traffic arrivals, that is, each sequence point of the time series is composed of the total number of traffic arrivals, which is called the sampling time step; Predict the input length of the training set\test set and the output length of the training set\test set; The actual start time to end time corresponding to each training set input sequence is called the input sequence corresponding time; The corresponding time of the input sequence is equal to the product of the sampling time step and the input length of the training set; Step 1-3: Build a classifier for all possible services. The features required to build a classifier include: The number of packets arriving within the corresponding time of the input sequence; The frequency of packet direction switching within the corresponding time of the input sequence; The protocol type with the largest number of arrivals within the corresponding time of the input sequence; The protocol type with the highest number of arrivals within the corresponding time of the input sequence; The full packet length of the input sequence corresponding to the time is the packet length with the most occurrences; The average packet length in the corresponding time of the input sequence; For a period of time corresponding to the input sequence data, its classification label and the input of the prediction training set can be obtained at the same time, that is, its classification label can correspond to the predicted value; Step 1-4: When actual business arrives, the classification features of the statistical business in step 1-3 are input into the classifier; the business classification result x is output, and the category labels l1, l2...l are selected. n The one closest to x is taken as the classification result; the implementation methods of selecting the nearest label include: ·Calculate x and l1, l2...l n The Euclidean distance of and selects the label with the smallest distance; ·Calculate x and l1, l2...l n The Hamming distance of and select the label with the smallest distance; ·Calculate x and l1, l2...l n Manhattan distance and select the label with the smallest distance; ·Calculate x and l1, l2...l n The cosine similarity of and select the label with the largest similarity; Step 1-5: Select the predictor corresponding to the business according to the business classification result, and divide the input of the test set according to the sampling time step, the corresponding time of the input sequence and the time window method, and input it into the predictor corresponding to the business for prediction; Step 2: Enhance traffic prediction by service classification without knowing the service category to which each traffic sequence belongs; Step 2-1: Count all the traffic data obtained and determine the input and output parameters of the classifier and predictor, including the following aspects: Sampling time step; Predict the input length of the training set\test set and the output length of the training set\test set; Input sequence corresponding time; Step 2-2: Extract the classification training set based on the acquired data. The method is to sample the existing business data, randomly select the start time to calculate the corresponding business classification features within the corresponding time of the future input sequence. The selection of business classification features is the same as step 1-3; Step 2-3: Build a business classifier and select a clustering method to classify the classification training set sampling results of step 2-2. The classification algorithms include: K-Means clustering method; Hierarchical clustering method; Fuzzy C-means calculation method; Step 2-4: Classify the arriving services. When a service arrives, the classification features of the service are counted according to steps 1-3 and input into the classifier to obtain the current service category. Step 2-5: Extract the prediction training set based on the acquired data; the business data within the corresponding time of the input sequence can be used to generate classification labels, i.e., the input of the classification training set and the prediction training set, and the implementation method is as follows: · Filter the part of the sampled data in step 2-2 whose business classification results are the same as the current business category, and construct it according to the method of steps 1-2 based on this part of business data; Sampling existing business data, randomly selecting the start time to calculate the corresponding business classification features within the corresponding time of the future input sequence and inputting them into the business classifier, collecting business data with the same business category as the current business and constructing them according to steps 1-2; Step 2-6: Use the generated prediction training set to build a predictor, divide the input of the test set according to the time window method for the current arriving business, and use the constructed predictor to make predictions; Step 3: When the service category of each traffic sequence is known, traffic prediction is used to enhance the accuracy of service classification; Step 3-1: Count all possible services that may arrive and give each service a numerical label, denoted as l1, l2, ...l n , the digital label can be a one-dimensional number or a multi-dimensional vector, where n is the total number of services that may arrive; Step 3-2: Build predictors for all possible arriving services; Data collection is performed on each possible business to obtain training sets and test sets, which are then input into the learner for learning to form net1, net2...net n ; The features required to build a predictor include the following information: The time step unit for counting the total number of traffic arrivals, that is, each sequence point of the time series is composed of the total number of traffic arrivals, which is called the sampling time step; Predict the input length of the training set\test set and the output length of the training set\test set; The actual start time to end time of each training set input sequence is called the input sequence corresponding time; the input sequence corresponding time is equal to the product of the sampling time step and the training set input length; Step 3-3: Build a classifier for all possible services. The features required to build a classifier include: The number of packets arriving within the corresponding time of the input sequence; The frequency of packet direction switching within the corresponding time of the input sequence; The protocol type with the largest number of arrivals within the corresponding time of the input sequence; The protocol type with the highest number of arrivals within the corresponding time of the input sequence; The full packet length of the input sequence corresponding to the time is the packet length with the most occurrences; The average packet length in the corresponding time of the input sequence; For a period of time corresponding to the input sequence data, its classification label and the input of the prediction training set can be obtained at the same time, that is, its classification label can correspond to the predicted value; Step 3-4: Sample and predict the test set according to the sampling time step, the corresponding time of the input sequence and the time window method, and record the corresponding business category, input it into each predictor for prediction, and calculate the error E1, E2...E of each predictor according to the true value of the business sequence. n ; Step 3-5: Import the current incoming data into the classifier for classification, calculate the current classification result x and each category label l1, l2...l n The distances d1, d2, ...d n , the distance calculation algorithm is the same as steps 1-4; Step 3-6: Set the weight factor α, where α is a positive value; calculate the final result of the distance between the business category and each classifier label [d1+αE1, d2+αE2...d n +αE n ], select the business category corresponding to the minimum value as the final classification result, and compare it with the real business category; Step 3-7: Repeat steps 3-4 to 3-6 to calculate the average classification accuracy, and improve the classification accuracy by adjusting the value of the weight factor α and select the α with the highest classification accuracy; Step 3-8: Determine the weight factor α After that, the real business sequence is input into each predictor and the errors E1, E2, ... E of each predictor are calculated respectively. n , calculate the final classification result according to steps 3-5 and 3-6.

2. The method for joint service classification and traffic time series prediction in a wireless network according to claim 1, characterized in that: The implementation methods of constructing the learner in step 1-2 include statistical models, machine learning models, neural network models, and hybrid models.

3. The method for joint service classification and traffic time series prediction in a wireless network according to claim 1, characterized in that: In step 3-4, the error calculation method includes: Root mean square error; Mean absolute error.

4. The method for joint service classification and traffic time series prediction in a wireless network according to claim 1, characterized in that: In the steps 3-7, the method for calculating α is: ·Acquired from the experience value constructed based on the business mathematical model; Traverse α within a certain range and make a selection; Use machine learning to perform parameter optimization and select the best α.

5. A computer program, characterized in that The computer program enables a computer to execute the method according to any one of claims 1 to 4.

6. An electronic device, characterized in that: include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device executes the method as claimed in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

8. A chip, characterized in that: include: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes a method as claimed in any one of claims 1 to 4.

9. A computer program product, characterized in that The computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method according to any one of claims 1 to 4 is implemented.

Images