CN105517019A - Method for detecting LTE (Long Term Evolution) network performance by using integrated regression system - Google Patents

Abstract

The invention provides a method for detecting an LTE (Long Term Evolution) network performance by using an integrated regression system. The method comprises the following steps: step (1), collecting a sample data set; step (2), through an established model, using a reception index to predict an empty delay, and adopting a regression algorithm to respectively predict samples in a training set on a specific subset; step (3), deducing an error value of each sample in the training set and an error of the regression algorithm; step (4), using an analytical method to perform weighted processing on the error in the regression algorithm, and forming an integrated regression prediction model through weighted regression; and step (5), applying the integrated regression algorithm in step (4) to a test set, and detecting the accuracy of a model prediction relationship obtained based on the training set. Through the operation and promotion of the method for detecting the LTE network performance by using the integrated regression system provided by the invention, the empty delay of a user level can be analyzed and inferred, so that a mobile operator can recognize problem cells with a higher empty delay, thus the service quality of an LTE network can be improved by optimizing the problem cells.

Description

Integrated regression system is adopted to detect the method for LTE network performance

Technical field

The present invention relates to a kind of method detecting LTE network performance, especially relate to a kind of method adopting integrated regression system to detect LTE network performance.

Background technology

In recent years, along with the fast development of LTE wireless network, generation and the acquisition of communication data are also thereupon flourish.Except the store and management to this data flow, a larger challenge how to utilize these data serving communication network better.Therefore, by converting data to corresponding network index and carry out critic network performance and user experience quality (QoE) becoming final goal.By analyzing, the time delay of eating dishes without rice or wine due to higher value will affect network reception quality and increase network interferences, and this experiences to have on user awareness comparatively directly affects, and time delay of therefore eating dishes without rice or wine is as a core index.Generally, time delay of eating dishes without rice or wine adopts acquisition by soft, but software and hardware input cost is higher, does not possess all-round popularization condition.

For mobile communication business, most important time delay is end-to-end time delay, and namely for the transmitting-receiving two-end connected, packet produces from transmitting terminal, to the time delay that receiving terminal correctly receives.Different according to business model, end-to-end time delay can be divided into one way time delay and backhaul time delay, wherein one way time delay refers to that packet produces the time delay correctly arriving another one receiving terminal through wireless network from transmitting terminal, and backhaul time delay refers to that packet produces destination server from transmitting terminal and receives packet and return corresponding packet until transmitting terminal correctly receives the time delay of reply data bag.

Existing mobile communication is interpersonal communication mainly, and along with miniaturization and the intellectuality of hardware device, the high speed between following mobile communication more " people and thing " and " thing and thing " is connected to be applied.Machine communication (MachineTypeCommunication, MTC) service application scope widely, as portable medical, car networking, Smart Home, Industry Control, environmental monitoring etc. will promote MTC system application explosive growth, large number quipments is by access network, realize real " all things on earth is interconnected ", for mobile communication brings boundless vital force.Simultaneously, MTC system range of application also can bring new technological challenge to mobile communication widely, the business such as such as real-time cloud calculating, virtual reality, game on line, tele-medicine, intelligent transportation, intelligent grid, remote real_time control are more responsive to time delay, propose higher demand to time delay.

Therefore, need to work out a kind of detection method, analyze and infer the time delay of eating dishes without rice or wine of user class, thus make mobile operator identify the problem cells of higher time delay of eating dishes without rice or wine, and then improve LTE network service quality by optimization problem community.

Summary of the invention

The object of this invention is to provide a kind of method detecting LTE network performance with integrated regression system makes mobile operator identify the problem cells of higher time delay of eating dishes without rice or wine to solve, and then improves the problem of LTE network service quality by optimization problem community.

Technical solution of the present invention is:

Adopt integrated regression system to detect the method for LTE network performance, comprise the following steps:

(1) collection of sample data collection: the reception achievement data of LTE network and network each stage time delay of eating dishes without rice or wine is carried out to collection and formed sample data collection, and described sample data collection is divided into training set and test set;

Wherein, training set is for finding the projected relationship receiving index and eat dishes without rice or wine between time delay; Test set is for detecting the accuracy of the model prediction relation drawn based on training set;

(2) in the operating process of described training set, by the model set up, utilize and receive index prediction and to eat dishes without rice or wine time delay, adopt regression algorithm to predict the sample in training set in particular subset respectively;

(3) by comparing the predicted value that in above-mentioned steps (2), model obtains and time delay value of eating dishes without rice or wine really, the error amount of each sample and the error of regression algorithm in training set is derived;

Wherein, for regression algorithm J, the row error amount in training set is then called as error J;

(4) adopt analytical method that the error in regression algorithm is weighted process, form integrated regressive prediction model by weighted regression combination;

(5) the integrated regression algorithm in step (4) is applied to test set, detect the accuracy of the model prediction relation drawn based on training set.

Further, the reception index receiving achievement data in described step (1) comprises Reference Signal Received Power RSRP, Reference Signal Received Quality RSRQ, Signal to Interference plus Noise Ratio SINR, the Physical Resource Block PUSCH-PRB on Physical Uplink Shared Channel and the Physical Resource Block PDSCH-PRB on Physical Downlink Shared Channel.

Further, the regression algorithm that employing eight kinds is different in step (2) is predicted the sample in training set respectively in particular subset, linear regression respectively, second order polynomial regression, three rank polynomial regressions, ridge regression, LASSO returns, Elastic returns, and GAM returns and MARS returns;

Wherein, described linear regression provides following formula (1):

E (y)=β ₀+ β ₁x ₁+ ...+β _dx _d(1), in formula, E (y) represents predicted value, and y represents hypothesis and to eat dishes without rice or wine time delay, x ₁..., x _drepresentative receives index; In this model, response variable y Gaussian distributed, uses least square method directly to calculate and obtains corresponding fitting coefficient β ₀..., β _d;

Described second order polynomial regression algorithm first calculate each index set once, Quadratic Orthogonal multinomial, thus obtain 2D form, carry out models fitting with 2D+1 item parameter;

Described three rank polynomial regression algorithm, when the choosing of variable, the number of times of orthogonal polynomial is chosen from 1 time to 3 times, thus decreases the constraint of forecast model;

Described ridge regression algorithm provides following formula (2):

in formula, algorithm is by increasing penalty coefficient, to factor beta ₀..., β _dlimit, become contraction, thus find the minimum variance of least squares estimator, the theory that wherein parametric t proposes based on control forecasting value variance according to E.Cule and M.DeIorio is chosen automatically, and k represents a kth coefficient.The scope of k from 1,2,3 ..., d;

Described LASSO regression algorithm provides following formula (3):

constraint function in formula limits the absolute value sum of regression coefficient, removing constant coefficient, and t value is chosen automatically;

Described Elastic regression algorithm provides following formula (4):

constraint function in formula is the linear combination constraint function in described ridge regression algorithm and described LASSO regression algorithm being carried out to regularization, and wherein, α is that 1/2, t value is chosen automatically;

Described GAM regression algorithm provides following formula (5):

G (E (y))=β ₀+ f ₁(x ₁)+...+f _d(x _d) (5), in formula, g represents generalized linear Copula, f ₁..., f _drepresent the non-linear relationship between input variable, β ₀for constant term, x ₁..., x _dreceive from five the data obtained index;

Described MARS regression algorithm provides following formula (6):

X → max (0, x-C) orx → max (0, C-x); C ∈ R (6), in MARS returns, returns the linear combination being synthesized to a hinge function.

Further, in described step (2), the calculating on training set is all completed by successive Regression, by finding suitable weight.

Further, in step (4), utilize the row error obtained in step (3) to infer weight, utilize the method for weight optimization model to give following formula (7):

wherein w ₁... w ₂drawn by minimizing Weighted least square method and the total weight calculation of restriction.

Further, infer in described step (4) in the process of weight, give following formula (8) for constraints:

${\mathrm{\Σ}}_{j=1}^8{w}_j=1---\left(8\right).$

Further, in described step (5), the predicted value obtained in test set and known time delay value of truly eating dishes without rice or wine are compared, and give formula (9) and draw error rate ε _test;

${\mathrm{\ϵ}}_{test}=\frac{1}{n_{test}}{\mathrm{\Σ}}_{i=1}^{n_{test}}\left|{\hat{y}}_i-y_i\right|---\left(9\right),$ In order to confirm stability and the accuracy of the integrated regressive prediction model obtained in described step (4), needing to carry out various different comparison, first, the identical error rate being performed on training set, obtaining an error rate ε _trsinif, ε _testand ε _trsinbetween difference less, then show not occur over-fitting, this means the stability of integrated regressive prediction model; Secondly, test set calculates the predicted value of often kind of regression algorithm, makes it respectively obtain an error rate, by the ε these error rates and integrated recurrence obtained _testcompare, then can check out the accuracy of integrated regressive prediction model.

The invention has the beneficial effects as follows: adopted the integrated regression system in the method for integrated regression system detection LTE network performance by the present invention, the prediction of effectively deriving of cell RF index is utilized to eat dishes without rice or wine time delay, integrated regressive prediction model is not only stablized but also can accurately be predicted time delay of eating dishes without rice or wine, it can carry out the selection of calculating and model automatically, and therefore model possesses stronger adaptivity; By examples prove, the theory of this assessment models is accurately, and from the angle of user, integrated regression system is easier to obtain and analyze index of correlation data, thus the predicting the outcome of the time delay that obtains eating dishes without rice or wine.Promoted by the operation adopting integrated regression system to detect the method for LTE network performance, can analyze and infer the time delay of eating dishes without rice or wine of user class, thus make mobile operator identify the problem cells of higher time delay of eating dishes without rice or wine, and then improve LTE network service quality by optimization problem community.

Accompanying drawing explanation

Fig. 1 is that the operation workflow that the embodiment of the present invention adopts integrated regression system to detect the method for LTE network performance illustrates schematic diagram;

Fig. 2 gives the simulated data sets comparative illustration schematic diagram of the integrated regression system of the present invention and classical regression algorithm.

Embodiment

The preferred embodiments of the present invention are described in detail below in conjunction with accompanying drawing.

Embodiment provides a kind of method adopting integrated regression system to detect LTE network performance, and the enforcement of this detection method comprises (1) data source: receive cell RF performance index data set; (2) associate: associating of the time delay value of eating dishes without rice or wine of cell RF index and each element of data centralization; (2) predict: be newly worth by cell RF index and estimate time delay of eating dishes without rice or wine.Other implementation contents except Section 1: sample data: sample data is distributed according to training set and test set; Modeling process: process data transporting something containerized statistical regression algorithm, calculate often kind of regression algorithm weight, in conjunction with regression algorithm and weight, to obtain Aksu River anticipation function; Modelling verification: on fc-specific test FC collection, model result is verified, the then accuracy of known predicted value.

Embodiment

Adopt integrated regression system to detect the method for LTE network performance, as shown in Figure 1, comprising:

(1) collection of sample data collection: the reception achievement data of LTE network and network each stage time delay of eating dishes without rice or wine is carried out to collection and formed sample data collection, and described sample data collection is divided into training set and test set;

Wherein, training set is for finding the projected relationship receiving index and eat dishes without rice or wine between time delay; Test set is for detecting the accuracy of the model prediction relation drawn based on training set;

(2) in the operating process of described training set, by the model set up, utilize and receive index prediction and to eat dishes without rice or wine time delay, adopt regression algorithm to predict the sample in training set in particular subset respectively;

(3) by comparing the predicted value that in above-mentioned steps (2), model obtains and time delay value of eating dishes without rice or wine really, the error amount of each sample and the error of regression algorithm in training set is derived;

Wherein, for regression algorithm J, the row error amount in training set is then called as error J;

(4) adopt analytical method that the error in regression algorithm is weighted process, form integrated regressive prediction model by weighted regression combination;

(5) the integrated regression algorithm in step (4) is applied to test set, detect the accuracy of the model prediction relation drawn based on training set.

Further, the reception index receiving achievement data in described step (1) comprises Reference Signal Received Power RSRP, Reference Signal Received Quality RSRQ, Signal to Interference plus Noise Ratio SINR, the Physical Resource Block PUSCH-PRB on Physical Uplink Shared Channel and the Physical Resource Block PDSCH-PRB on Physical Downlink Shared Channel.

Further, the regression algorithm that employing eight kinds is different in step (2) is predicted the sample in training set respectively in particular subset, linear regression respectively, second order polynomial regression, three rank polynomial regressions, ridge regression, LASSO returns, Elastic returns, and GAM returns and MARS returns;

Wherein, described linear regression provides following formula (1):

E (y)=β ₀+ β ₁x ₁+ ...+β _dx _d(1) in formula, E (y) represents predicted value, and y represents hypothesis and to eat dishes without rice or wine time delay, x ₁..., x _drepresentative receives index; In this model, response variable y Gaussian distributed, uses least square method directly to calculate and obtains corresponding fitting coefficient β ₀..., β _d;

Described second order polynomial regression algorithm first calculate each index set once, Quadratic Orthogonal multinomial, thus obtain 2D form, carry out models fitting with 2D+1 item parameter;

Described three rank polynomial regression algorithm, when the choosing of variable, the number of times of orthogonal polynomial is chosen from 1 time to 3 times, thus decreases the constraint of forecast model;

Described ridge regression algorithm provides following formula (2):

in formula, algorithm is by increasing penalty coefficient, to factor beta ₀..., β _dlimit, become contraction, thus find the minimum variance of least squares estimator, the theory that wherein parametric t proposes based on control forecasting value variance according to E.Cule and M.DeIorio is chosen automatically, k represents a kth coefficient, the scope of k from 1,2,3 ..., d;

Described LASSO regression algorithm provides following formula (3):

constraint function in formula limits the absolute value sum of regression coefficient, removing constant coefficient, and t value is chosen automatically;

Described Elastic regression algorithm provides following formula (4):

${\mathrm{\α\Σ}}_{k=1}^d{\mathrm{\β}}_k^2+(1-\mathrm{\α}){\mathrm{\Σ}}_{k=1}^d\left|{\mathrm{\β}}_k\right|\≤t---\left(4\right),$ Constraint function in formula is the linear combination constraint function in described ridge regression algorithm and described LASSO regression algorithm being carried out to regularization, and wherein, α is that 1/2, t value is chosen automatically;

Described GAM regression algorithm provides following formula (5):

G (E (y))=β ₀+ f ₁(x ₁)+...+f _d( _xd) (5), in formula, g represents generalized linear Copula, f ₁..., f _drepresent the non-linear relationship between input variable, β ₀for constant term, x ₁..., x _dreceive from five the data obtained index;

Described MARS regression algorithm provides following formula (6):

X → max (0, x-C) orx → max (0, C-x); C ∈ R (6), in MARS returns, returns the linear combination being synthesized to a hinge function.

Further, in described step (2), the calculating on training set is all completed by successive Regression, by finding suitable weight.

Further, in step (4), utilize the row error obtained in step (3) to infer weight, utilize the method for weight optimization model to give following formula (7):

wherein w ₁... w _sdrawn by minimizing Weighted least square method and the total weight calculation of restriction.

Further, infer in described step (4) in the process of weight, give following formula (8) for constraints:

${\mathrm{\Σ}}_{j=1}^8{w}_j=1---\left(8\right).$

Further, in described step (5), the predicted value obtained in test set and known time delay value of truly eating dishes without rice or wine are compared, and give formula (9) and draw error rate ε _test:

${\mathrm{\ϵ}}_{test}=\frac{1}{n_{test}}{\mathrm{\Σ}}_{i=1}^{n_{test}}\left|{\hat{y}}_i-y_i\right|---\left(9\right),$ In order to confirm stability and the accuracy of the integrated regressive prediction model obtained in described step (4), needing to carry out various different comparison, first, the identical error rate being performed on training set, obtaining an error rate ε _trsinif, ε _testand ε _trsinbetween difference less, then show not occur over-fitting, this means the stability of integrated regressive prediction model; Secondly, test set calculates the predicted value of often kind of regression algorithm, makes it respectively obtain an error rate, by the ε these error rates and integrated recurrence obtained _testcompare, then can check out the accuracy of integrated regressive prediction model.

Fig. 1 gives embodiment and adopts integrated regression system to detect the operation workflow explanation schematic diagram of the method for LTE network performance.In more detail, the object that the integrated regression system of this employing detects the method for LTE network performance is that conventional receiver index prediction by describing in detail is below eated dishes without rice or wine the approximation of time delay, but be more critically, present slight non-linear between time delay of eating dishes without rice or wine and conventional receiver index, there is no direct correlation.Therefore, algorithm model can not process the index received simply by linear regression.Multiple regression algorithm, when predicting large data, is attached in final weight estimation, just can increases system robustness, and obtain predicted value accurately by this system (the integrated regressive prediction model in the present invention).

Legacy system can carry out summary by the flow chart shown in Fig. 1 and describe.First complete and receive achievement data and network each stage and to eat dishes without rice or wine the collection of time delay.This sample data collection is divided into training set and test set by model.Wherein the data of 70% be training set, 30% data be test set.

Secondly, training set is for finding the projected relationship receiving index and eat dishes without rice or wine between time delay.In the operating process of training set, by the model set up, utilize and receive index prediction and to eat dishes without rice or wine time delay.Specifically, be exactly the different regression algorithm of use eight kinds, in particular subset, the sample in training set predicted respectively.(by staying a cross-validation method, will describe in detail in ensuing paragraph).

The predicted value obtained by comparison model and time delay value of eating dishes without rice or wine really, thus derive the error amount of each sample and the error of often kind of regression algorithm in training set.For regression algorithm J, the row error amount in training set is then called as " error J ".

Afterwards, take a kind of analytical method that regression algorithm is weighted process (be called least square method, will introduce in detail in ensuing paragraph) according to error 1 ~ 8.Algorithm after being combined by weighted regression just constitutes our integrated regressive prediction model.

Finally, integrated regression algorithm is applied to test set.Detect the accuracy of the model prediction relation drawn based on training set with it, and then improve the forecasting accuracy of whole model.According to Mathematical Modeling proof rule, model result can carry out matching accurately at training set and checking collection, and so not think the situation that there is overfitting, then this model is accurately.

The present embodiment, except describing time delay of eating dishes without rice or wine, also relates to the reception index that other five are used for weighing the quality of reception and MPS process.They and eat dishes without rice or wine between time delay, to there is certain positive and negative dependency relation.These five indexs are specifically: Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), Signal to Interference plus Noise Ratio (SINR), the Physical Resource Block (PUSCH-PRB) on Physical Uplink Shared Channel and the Physical Resource Block (PDSCH-PRB) on Physical Downlink Shared Channel.RSRP represents on the frequency band measured, the linear average of the power of the Resource Unit of carrying community own reference signal, i.e. average signal strength.Measured the power of Received signal strength by the mean value of some reference signals, therefore can know whether user easily accesses this community by it.The ratio of what RSRQ was corresponding is received signal strength and total bandwidth power, wherein, total bandwidth power comprises Serving cell power, interference power and noise.It is that the level of noise and signal disturbing is by quantifying to assess the quality of received signal.SINR is the ratio between signal power and interference power, noise summation.PUSCH-PRB (PDSCH-PRB) represents the PRB quantity distributing to up channel (down channel), and user is an important reference at the numerical value of local resource consumption.

Sample data is divided into training set and test set, when the training set based on sample data carries out regression training, first need to understand in native system involved eight kinds of different regression algorithms.As mentioned above, these eight kinds of regression algorithms have used different statistical methods and statistic algorithm, because this ensure that the accuracy of integrated regression system.These eight kinds of regression algorithms are: linear regression, second order polynomial regression, three rank polynomial regressions, ridge regression, and LASSO returns, and Elastic returns, and GAM returns and MARS returns.

In classical linear regression, suppose to eat dishes without rice or wine time delay (hereinafter referred to as y) (being called x with reception index ₁..., x _d) between pass be linear.Using formula 1 represents relation between the two, and wherein E (y) represents predicted value.In this model, response variable y Gaussian distributed, uses least square method directly to calculate and obtains corresponding fitting coefficient β ₀..., β _d.Take this regression algorithm, final predicted value will be subject to the impact of all reception indexs.

E(y)＝β ₀+β ₁x ₁+...+β _dx _d(1)

Second order polynomial regression algorithm is comparatively close to classical linear regression.But it is not directly select to receive index x ₁..., x _d, but first calculate each index set once, Quadratic Orthogonal multinomial, thus obtain 2D form.After having made this change, carry out models fitting with 2D+1 (the 2D form of index and constant term) item parameter, instead of the D+1 item parameter in classical linear model.This fit approach operation is upper the same with least square method before.

Three rank polynomial regression algorithm, model is similar to second order polynomial, but when the choosing of variable, the number of times of orthogonal polynomial is chosen from 1 time to 3 times, thus decreases the constraint of forecast model.

It is similar that ridge regression algorithm and formula 1 describe, but unlike, this algorithm passes through to increase penalty coefficient, to factor beta ₀..., β _dcarry out limiting (becoming contraction).Precisely, this algorithm by limiting the quadratic sum of coefficient, thus finds the minimum variance of least squares estimator (being similar to classical linear regression), as shown in equation 2.The theory that wherein parametric t proposes based on control forecasting value variance according to E.Cule and M.DeIorio is chosen automatically.

${\mathrm{\Σ}}_{k=1}^d{\mathrm{\β}}_k^2\≤t---\left(2\right)$

LASSO return and ridge regression a bit similar, but wherein constraint function difference.The constraint function that LASSO returns limits the absolute value sum (removing constant coefficient) of regression coefficient, as shown in Equation 3.Theory analysis shows that this is more strong constraint for coefficient, even can produce some strict null coefficient.Therefore, this algorithm by making alternatively to carry out coefficient of reduction, thus achieves the object that index set simplifies.In ridge regression, t value is chosen automatically.

${\mathrm{\Σ}}_{k=1}^d\left|{\mathrm{\β}}_k\right|\≤t---\left(3\right)$

It is the another kind of algorithm by the coefficient of reduction that carries out in returning at ridge regression and LASSO compromising that Elastic returns.Here constraint function is the linear combination constraint function in ridge regression and LASSO recurrence being carried out to regularization, as shown in Equation 4.Wherein, α is that 1/2, t value is chosen automatically.

${\mathrm{\α\Σ}}_{k=1}^d{\mathrm{\β}}_k^2+(1-\mathrm{\α}){\mathrm{\Σ}}_{k=1}^d\left|{\mathrm{\β}}_k\right|\≤t---\left(4\right)$

In GAM returns, eat dishes without rice or wine time delay value y and index input variable x ₁..., x _dbetween relation as shown in Equation 5, wherein suppose y from ED~* class.On the left side of formula, G represents generalized linear Copula.F on the right of formula ₁..., f _drepresent the non-linear relationship between input variable, β ₀for constant term.Function f ₁a kind of nonparametric backfitting algorithm can be considered to.This algorithm uses the mode of iteration, utilizes cubic spline function (at least there is the cube of certain index) to carry out approximate description.In this article, y represents and eats dishes without rice or wine time delay and suppose Gaussian distributed (belonging to exponential family), g=id, x ₁..., x _dreceive from five the data obtained index.

g(E(y))＝β ₀+f ₁(x ₁)+...+f _d(x _d)(5)

In MARS returns, return the linear combination being synthesized to a hinge function.Hinge function, as in formula 6 define, be by the overall nonlinear model of self non-linear release.As a whole, whole space is divided into the polynomial subspace of each response.Return operation to be separated these spaces and performing, need two steps.First, by reducing the error of residual sum of squares (RSS), the basic function that makes new advances of forward calculation and hinge function in an iterative manner.Next, then by backward deletion minimum contribution item, model is revised.Final step is then reduce over matching.

x→max(0，x-C)orx→max(0，C-x)；C∈R(6)

Calculating on training set is all completed by successive Regression.By finding suitable weight, thus often kind of regression combination is become integrated recurrence.Next will be described this process.

In each regression algorithm J, element i each on training set is adopted and stays a cross-validation method, thus obtain predicted value in order to obtain this predicted value, then to carry out the operation of regression algorithm j to all elements except i.So obtain one to be obtained by element i anticipation function.When eating dishes without rice or wine, Yanzhong all elements is all known, variance then by calculating.Therefore, for each regression algorithm J, column vector error is then defined as: ε _j:=(ε _ij) ' _i(wherein ' represent transposition).

Next step utilizes these row errors to infer weight.W ₁... w ₈drawn by minimizing Weighted least square method and the total weight calculation of restriction.Wherein, formula 7 defines the method utilizing weight optimization model, and formula 8 defines constraints.Here || || ₂represent Euclid norm.

$\underset{w_1\mathrm{...}{w}_s}{\arg \min }\left|\right|{\mathrm{\Σ}}_{j=1}^8{w}_j{\mathrm{\ϵ}}_j|{|}_2---\left(7\right)$

${\mathrm{\Σ}}_{j=1}^8{w}_j=1---\left(8\right)$

After obtaining weight, still need to understand the regression function quoted in often kind of prediction and often kind of regression algorithm.In order to realize this point, all elements on training set is all performed by regression algorithm j, thus obtain anticipation function P _j.Finally, integrated recurrence is then defined as P _jlinear weighted combination.

Once integrated recurrence is drawn by derivation, then its result is checked in test set.In checking process, the predicted value obtained in test set and known time delay value of truly eating dishes without rice or wine are compared.The average error rate defined in formula 9 is calculated error rate ε _test.In order to confirm stability and the accuracy of integrated regression algorithm, need to carry out various different comparison.First, the identical error rate is performed on training set, obtain an error rate ε _trsin.If ε _testand ε _trsinbetween difference less, then show not occur over-fitting, this means the stability of algorithm.Secondly, test set calculates the predicted value of often kind of regression algorithm, makes it respectively obtain an error rate.By the ε that these error rates and integrated recurrence are obtained _testcompare, then can check out the accuracy of algorithm.

${\mathrm{\ϵ}}_{test}=\frac{1}{n_{test}}{\mathrm{\Σ}}_{i=1}^{n_{test}}\left|{\hat{y}}_i-y_i\right|---\left(9\right)$

Fig. 2 gives a sample of analogue data, according to said process, calculates associated weight and derive integrated regression algorithm by analytic approach, calculating average error rate at training set and test set.In this example, the error increment of training set and test set is lower, shows that model does not occur overfitting.Therefore known, integrated regression algorithm is for other regression algorithm, and the result drawn is more accurate, and practicality is stronger.In addition, relative to other all algorithms (or even three rank polynomial regression algorithm), the stability of integrated recurrence is also stronger.Therefore, preferred regression algorithm is exactly integrated recurrence here.

In general, the prediction of effectively deriving of cell RF index is utilized to eat dishes without rice or wine time delay by integrated regression system.Integrated regression model is not only stablized but also can accurately be predicted time delay of eating dishes without rice or wine, and it can carry out the selection of calculating and model automatically, and therefore model possesses stronger adaptivity.By examples prove, the theory of this assessment models is accurately.From the angle of user, native system is easier to obtain and analyze index of correlation data, thus the predicting the outcome of the time delay that obtains eating dishes without rice or wine.

Claims (7)

1. adopt integrated regression system to detect a method for LTE network performance, comprise the following steps:

(1) collection of sample data collection: the reception achievement data of LTE network and network each stage time delay of eating dishes without rice or wine is carried out to collection and formed sample data collection, and described sample data collection is divided into training set and test set;

Wherein, training set is for finding the projected relationship receiving index and eat dishes without rice or wine between time delay; Test set is for detecting the accuracy of the model prediction relation drawn based on training set;

(2) in the operating process of described training set, by the model set up, utilize and receive index prediction and to eat dishes without rice or wine time delay, adopt regression algorithm to predict the sample in training set in particular subset respectively;

(3) by comparing the predicted value that in above-mentioned steps (2), model obtains and time delay value of eating dishes without rice or wine really, the error amount of each sample and the error of regression algorithm in training set is derived;

Wherein, for regression algorithm J, the row error amount in training set is then called as error J;

(4) adopt analytical method that the error in regression algorithm is weighted process, form integrated regressive prediction model by weighted regression combination;

(5) the integrated regression algorithm in step (4) is applied to test set, detect the accuracy of the model prediction relation drawn based on training set.

2. the method adopting integrated regression system to detect LTE network performance as claimed in claim 1, it is characterized in that, the reception index receiving achievement data in described step (1) comprises Reference Signal Received Power RSRP, Reference Signal Received Quality RSRQ, Signal to Interference plus Noise Ratio SINR, the Physical Resource Block PUSCH-PRB on Physical Uplink Shared Channel and the Physical Resource Block PDSCH-PRB on Physical Downlink Shared Channel.

3. the method adopting integrated regression system to detect LTE network performance as claimed in claim 2, it is characterized in that, the regression algorithm that employing eight kinds is different in step (2) is predicted the sample in training set respectively in particular subset, linear regression respectively, second order polynomial regression, three rank polynomial regressions, ridge regression, LASSO returns, and Elastic returns, and GAM returns and MARS returns;

Wherein, described linear regression provides following formula (1):

E (y)=β ₀+ β ₁x ₁+ ...+β _dx _d(1), in formula, E (y) represents predicted value, and y represents hypothesis and to eat dishes without rice or wine time delay, x ₁..., x _drepresentative receives index; In this model, response variable y Gaussian distributed, uses least square method directly to calculate and obtains corresponding fitting coefficient β _o..., β _d;

Described second order polynomial regression algorithm first calculate each index set once, Quadratic Orthogonal multinomial, thus obtain 2D form, carry out models fitting with 2D+1 item parameter;

Described three rank polynomial regression algorithm, when the choosing of variable, the number of times of orthogonal polynomial is chosen from 1 time to 3 times, thus decreases the constraint of forecast model;

Described ridge regression algorithm provides following formula (2):

in formula, algorithm is by increasing penalty coefficient, to factor beta ₀..., β _dlimit, become contraction, thus find the minimum variance of least squares estimator, the theory that wherein parametric t proposes based on control forecasting value variance according to E.Cule and M.DeIorio is chosen automatically, k represents a kth coefficient, the scope of k from 1,2,3 ..., d;

Described LASSO regression algorithm provides following formula (3):

constraint function in formula limits the absolute value sum of regression coefficient, removing constant coefficient, and t value is chosen automatically;

Described Elastic regression algorithm provides following formula (4):

${\mathrm{\α\Σ}}_{k-1}^a{\mathrm{\β}}_k^2+(1-\mathrm{\α}){\mathrm{\Σ}}_{k-1}^d\left|{\mathrm{\β}}_k\right|\≤t---\left(4\right),$ Constraint function in formula is the linear combination constraint function in described ridge regression algorithm and described LASSO regression algorithm being carried out to regularization, and wherein, α is that 1/2, t value is chosen automatically;

Described GAM regression algorithm provides following formula (5):

G (E (y))=β ₀+ f ₁(x ₁)+...+f _d(x _d) (5), in formula, g represents that generalized linear contacts

Function, f ₁..., f _drepresent the non-linear relationship between input variable, β ₀for constant term, x ₁..., x _dreceive from five the data obtained index;

Described MARS regression algorithm provides following formula (6):

X → max (0, x-C) orx → max (0, C-x) C ∈ R (6), in MARS returns, returns the linear combination being synthesized to a hinge function.

4. the method adopting integrated regression system to detect LTE network performance as claimed in claim 3, it is characterized in that, in described step (2), the calculating on training set is all completed by successive Regression, by finding suitable weight.

5. the method adopting integrated regression system to detect LTE network performance as claimed in claim 4, it is characterized in that, in step (4), utilize the row error obtained in step (3) to infer weight, utilize the method for weight optimization model to give following formula (7):

wherein w ₁... w ₂drawn by minimizing Weighted least square method and the total weight calculation of restriction.

6. the method adopting integrated regression system to detect LTE network performance as claimed in claim 5, is characterized in that, infer in the process of weight, give following formula (8) for constraints in described step (4):

${\mathrm{\Σ}}_{j=1}^8{w}_j=1---\left(8\right).$

7. the method adopting integrated regression system to detect LTE network performance as claimed in claim 5, it is characterized in that, in described step (5), the predicted value obtained in test set and known time delay value of truly eating dishes without rice or wine are compared, and gives formula (9) and draw error rate ε _test:

${\mathrm{\ϵ}}_{test}=\frac{1}{n_{test}}{\mathrm{\Σ}}_{i=1}^{n_{test}}|{\hat{y}}_i-y_i|---\left(9\right),$ In order to confirm stability and the accuracy of the integrated regressive prediction model obtained in described step (4), needing to carry out various different comparison, first, the identical error rate being performed on training set, obtaining an error rate ε _trainif, ε _testand ε _treinbetween difference less, then show not occur over-fitting, this means the stability of integrated regressive prediction model; Secondly, test set calculates the predicted value of often kind of regression algorithm, makes it respectively obtain an error rate, by the ε these error rates and integrated recurrence obtained _testcompare, then can check out the accuracy of integrated regressive prediction model.