Summary of the invention
In order to solve, traditional power telecom network is unobvious to those signs, there are the performance informations of hidden danger to lack effectively
This realistic problem of diagnostic means, the performance parameter for the characterization network operation state that the present invention is provided from specialized network management system
Start with, based on Probability Theory and Math Statistics theory and according to Time-Series analysis basic theories and method, is become by constructing network
Gesture prediction model, energy Accurate Prediction go out the operating status of communication network, to be the active maintenance mould based on equipment state overhauling
Formula provides technical support.
The present invention solves the principles of science based on its technical problem:
Time series analysis (Time Series Analysis) is disclosed by the correlativity of analysis different moments variable
Its structure relevant to time series and rule are grasped inside data to recognize the inherent characteristic of generation time sequential system
System contacts rule with external, goes prediction and control future value from the past value of system.
If one group of random variable values, the sequence constituted is known as random sequence, with indicating.If subscript is integer variable,
Increment at the time of it represents time interval, such as t moment, the t days, the t times, this random sequence is referred to as time series.
The numerical value of a certain characteristic quantity can be obtained by the time sequence of this feature numerical quantity by certain time interval arrangement
Column.The time series variation of the optical transport network performance parameter monitored in reality is affected by many factors:
1) because accidental, indecisive enchancement factor influence shows its random fluctuation and scrambling.Such as
Optical transport network performance parameter in monitoring process by certain random disturbances so that the tables of data of its characteristic quantity reveal certain with
The fluctuation of machine.
2) become due to being made its variation show certain by certain fixed factors or period sexual factor, conclusive influence
Gesture and certain regularity.When optical transport network is influenced by certain incipient fault in monitoring process, performance parameter numerical value
It would generally show certain trend persistently risen or fallen.
3) it is influenced again by certain periodic service conditions in monitoring process, numerical value can show periodically again
Regular fluctuation.It is influenced again by certain periodic service conditions in monitoring process, numerical value can show periodically again
Regular fluctuation.
The purpose for introducing Time Series method herein is exactly want to influence characteristic quantity numerical value change by analyzing and distinguishing
Factor, and analyze respectively its clock synchronization ask sequence change rule, with disclose because becoming for a long time caused by communication equipment incipient fault
Gesture changing rule, and predict its future developing trend, technical support is provided to carry out communication network status maintenance.
To solve the above-mentioned problems, the technical solution used in the present invention is:
1, a kind of optical transport network trend forecasting method based on Time-Series analysis, it is characterised in that: specifically include following step
It is rapid:
Step 1: network management performance choice of parameters, according to authoritative national standard and professional standard, the experimental data at scene, a large amount of
Documents and materials and expert suggestion, selection is able to reflect the network management performance parameter of network operation state;
Step 2: network management performance data obtains, and obtains specified network management performance by the northbound interface of optical transmission device network management
Data;
Step 3: time series is formed, during the sampling period uninterrupted sampling network management performance data, by between the regular hour
Every the time series for being arranged in certain performance parameter feature numerical quantity;
Step 4: Time Series, by the analysis to timed sample sequence, decomposite trend term in time series,
Periodic term and random entry;
Step 5: subitem predictor calculation is distinguished for three kinds of different types of decomposition items according to respective prediction model
Estimate predicted value, trend term decomposes to obtain by sliding average algorithm, and periodic term is calculated by multicycle superposed average method, then
The prediction of subsequent cycle data is carried out using continuation method, random entry obtains the prediction number of random entry using ARMA (p, g) model
According to;
Step 6: final predictor calculation calculates final predicted value, and and true value according to time series addition model
Carry out cross validation.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: in step 1,
It include that background block misses on the influential performance parameter of the quality of service of communication network operating status and carrying in optical transmission device
Code, errored block, errored seconds, the bit error rate, unavailable second, Severely Errored Second, continuius severely errored second, FEC corrected bytes count,
Frame count that FEC cannot be corrected, CRC check mistake, Pointer Justification Count, Protection Switching Count PSC, operating temperature, input optical power,
Output optical power, laser bias current, wherein the bit error rate and optical power are measure communication network signal transmission quality important
Index.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: in step 3,
The formation of time series: setting one group of random variable values, and the sequence constituted is known as random sequence, uses Xt(t=1,2...n) table
Show, if subscript is integer variable, increment at the time of it represents time interval, such as t moment, the t days, the t times claim this
Random sequence is time series.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: in step 4, when
Between the detailed analysis process decomposed of sequence: according to modern Practical Statistic correlation theory, any time sequence is by reasonable letter
Transformation of variables, also regarded as by trend term, periodic term, random entry three parts are formed by stacking, the optical transport monitored in reality
The time series variation of network performance parameter is affected by many factors:
1) because accidental, indecisive enchancement factor influence shows its random fluctuation and scrambling, light is passed
Defeated network performance parameter is in monitoring process by certain random disturbances, so as to reveal certain random for the tables of data of its characteristic quantity
Fluctuation;
2) due to being influenced that its variation is made to show certain regularity by certain period sexual factors, in monitoring process
It is influenced again by certain periodic service conditions, numerical value can show periodic regular fluctuation;
3) due to being influenced by certain fixed factors, time series is in certain trend, when optical transport in monitoring process
When network is influenced by certain incipient fault, what performance parameter numerical value would generally show that certain persistently rises or falls becomes
Gesture;
According to influence communication network status monitoring quantity time series factor, the variation of time series substantially can be analyzed to
Lower three kinds of forms:
(1) tendency fluctuation: refer to that phenomenon is influenced to change over time towards certain orientation by fixed factor and show continually and steadily
Ground up and down or stable trend;
(2) cyclical variations: refer to that phenomenon is influenced the cyclic fluctuation shown by certain fixed cycle by period sexual factor, together
Sample, cyclical swing, which is also due to certain factors, causes data that cyclic fluctuation is presented;
(3) random fluctuation: refer to the irregular fluctuation that phenomenon is influenced by accidentalia and is shown;
The top priority of time series analysis be by analyzing collected performance parameter, will be in time series
Trend term, periodic term and random entry, which decomposite, to be come, and establishes different regression models respectively to three kinds of different type items of decomposition, and
Estimated by own intellectual energy supplemental characteristic, it is final to realize the trend prediction based on time series.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: right in step 5
It is predicted in trend term, is estimated using multiple linear regression model and predicted, the general type of multiple linear regression model are as follows:
Y=μ0+μ1x1+…+μmxm+ε (2)
Wherein μ0,μ1......μmIt is m+1 unknown parameter, referred to as regression coefficient, y is known as explained variable, and x1,
x2......xmIt is the general variances that m is collected, i.e. explanatory variable, as m=1, as Linear Regression Model in One Unknown;m≥
It is then multiple linear regression model when 2, ε indicates random error;
For the N group monitoring data that we obtain, i-th group can be indicated are as follows: yi=xi1, xi2.....xim, wherein i=1,
2,3 ... n, then linear regression model (LRM) may be expressed as:
Since there are residual error e between actual value and estimated valuei=yi-(μ0+μ1xi1+…+μmxim+εi), utilize least square
The estimation technique improves the precision of estimated value so that the quadratic sum of residual error is minimum:
Σei 2=Σ (yi-(μ0+μ1xi1+…+μmxim+εi))2 (4)
Its necessary condition are as follows:
It is computed, formula (4) can convert are as follows:
To the deviation of estimated value estimated value and actual value be therefore,By left and right simultaneously multiplied by x ', can be obtainedAnd x ' e=0,
Therefore:
Estimated value is calculated according to collected performance parameter sampleAnd verified by actual value, to obtain
Final regression equation:
Due toIt is the unbiased esti-mator of E (y), thus it is availableAs final predicted value.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: right in step 5
Periodic term StIt is decomposed, the data point on same phase just constitutes sequence.
If periodic term sequence is { T (t), t=1,2 ... N }, the data point of cycle T=h, same phase constitutes same row,
Then former sequence can transform to following matrix:
Above-mentioned matrix (8) is split into the subsequence of h out of phase, by column vector so as to complete the weight in the period
Structure models each subsequence, by changing rule of the simulation same phase in different cycles, can predict this phase
Data value of the position in next period, so that multi-step prediction is converted into Single-step Prediction, periodic term is obtained by Single moving average method
It arrives, i.e. multicycle observed value superposed average, using this mean value as the predicted value of next phase.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: in step 5, with
The prediction of machine item, is modeled and is predicted using the Time Series Analysis Method of autoregressive moving-average model, using autoregressive moving average
Model abbreviation arma modeling, ARMA, as long as determining limited parameter, can determine model as limited parameter model completely, benefit
The predicted value of random entry subsequent time is calculated with arma modeling mainly according to the p history value and q error before its t moment
Value is linearly calculated, and ARMA (p, g) model specifically can be described as:
V (t+1) is the predicted value of subsequent time random entry in formula (9),Be undetermined coefficient not equal to zero with θ, p and
Q is respectively arma modeling Autoregressive and moving average order;V (t+1-i) is the random entry measured value before t moment, ε (t
+ 1-j) be t moment before random entry predict error term, therefore, in random entry prediction model most importantly determine model parameter p
With the value of order q,
For the value of order q, there is nearly 95% numerical value to all fall within 2 after the auto-correlation coefficient of time series is in certain d rank
Within the scope of times standard deviation, and the wave process for decaying to small value of auto-correlation coefficient, then it is believed that model order d value thus,
The value of model parameter p is estimated using least square regression, obtained residual by acquiring N number of sample data
Difference are as follows:
Wherein, actual value when p=1,2 ..., N, V (t+1) are t+1, residual error e are the deviation of actual value and predicted value,
N is sample size, and according to the least square regression that formula (4) trend term is predicted, its solution, which can be obtained, is
After the value of model parameter p and order q determines, then random entry is predicted according to historical data.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: step 6: final
Predictor calculation is independent from each other due to influencing the factor of performance parameter characteristic quantity of communications network monitors, obtains
Premeasuring should be the sum of three kinds of factors influences, and the addition model of time series analysis can be used:
Xt=Mt+St+It (1)
Wherein: XtFor primitive data item, MtFor trend term, StFor periodic term, and meet St+d=St,D is indicated
Time interval number, ItFor random entry.
A kind of optical transport network trend forecasting method based on Time-Series analysis above-mentioned, it is characterised in that: to being calculated
Least-squares linear regression estimated valueCarry out it is necessary inspection and evaluation, that is, examine in aggregate level dependent variable and from
It whether there is linear relationship between variable, specifically include: the measurement of regression equation fitting degree, standard error estimate, recurrence side
The hypothesis testing of journey, the hypothesis testing of regression coefficient and multicollinearity differentiate.
Advantageous effects of the invention: the present invention is based on the optical transport network trend forecasting method of Time-Series analysis, it can
Technical support and judgment basis are provided for communication network status maintenance.According to time series analysis theory, from professional equipment network management
The performance parameter of provided characterization network operation state is started with, and based on Probability Theory and Math Statistics theory, establishes base
In the network performance characteristic quantity trend prediction model that time series variation decomposes, and pass through the performance parameter data pair of actual acquisition
Decomposition model and algorithm are verified.The method of the present invention can find the Hidden fault of communication equipment in advance, reduce logical
Believe equipment repair and maintenance expense, improves communication network safety in operation and reliability.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.Following embodiment is only used for clearly illustrating the present invention
Technical solution, and not intended to limit the protection scope of the present invention.
As shown in Figure 1, a kind of optical transport network trend forecasting method based on Time-Series analysis, it is characterised in that: specific packet
Include following steps:
Step 1: network management performance choice of parameters, according to authoritative national standard and professional standard, the experimental data at scene, a large amount of
Documents and materials and expert suggestion, selection is able to reflect the network management performance parameter of network operation state.
Step 2: network management performance data obtains, and obtains specified network management performance by the northbound interface of optical transmission device network management
Data;
Step 3: time series is formed, during the sampling period uninterrupted sampling network management performance data, by between the regular hour
Every the time series for being arranged in certain performance parameter feature numerical quantity;
Step 4: Time Series, by the analysis to timed sample sequence, decomposite trend term in time series,
Periodic term and random entry;
Step 5: subitem predictor calculation is distinguished for three kinds of different types of decomposition items according to respective prediction model
Estimate predicted value, trend term decomposes to obtain by sliding average algorithm, and periodic term is calculated by multicycle superposed average method, then
The prediction of subsequent cycle data is carried out using continuation method, random entry obtains the prediction number of random entry using ARMA (p, g) model
According to;
Step 6: final predictor calculation calculates final predicted value, and and true value according to time series addition model
Carry out cross validation.
Reckoning result of the present invention to trend prediction model:
According to influence communication network status monitoring quantity time series factor, the variation of time series substantially can be analyzed to
Lower three kinds of forms:
(1) tendency fluctuation: refer to that phenomenon is influenced to change over time towards certain orientation by fixed factor and show continually and steadily
Ground up and down or stable trend.
(2) cyclical variations: refer to that phenomenon is influenced the cyclic fluctuation shown by certain fixed cycle by period sexual factor.Together
Sample, cyclical swing, which is also due to certain factors, causes data that cyclic fluctuation is presented.
(3) random fluctuation: refer to the irregular fluctuation that phenomenon is influenced by accidentalia and is shown.
The top priority of time series analysis be by analyzing collected performance parameter, will be in time series
Trend term, periodic term and random entry, which decomposite, to be come.Different regression models is established respectively to three kinds of different type items of decomposition, and
Estimated by own intellectual energy supplemental characteristic, it is final to realize the trend prediction based on time series.
Which kind of specifically mainly taken using the Fluctuation or variation feature of each factor in method analysis and evaluation and test time series
Certainly in the hypothesis to correlation between three kinds of variables.There are mainly two types of add for the relationship of each variable of usual time series
Method relationship and multiplication relationship, form addition model or multiplied model accordingly.
Factor due to influencing the performance parameter characteristic quantity of communications network monitors is independent from each other, and is obtained pre-
Measurement should be the sum of three kinds of factors influences, and the addition model of time series analysis can be used:
Xt=Mt+St+It (1)
Wherein: XtFor primitive data item, MtFor trend term, StFor periodic term, and meet St+d=St,ItFor with
Machine item.
1) trend term prediction model
For trend term, multiple linear regression model usually can be used and estimated and predicted.Multiple linear regression model
General type are as follows:
Y=μ0+μ1x1+…+μmxm+ε (2)
Wherein μ0,μ1......μmIt is m+1 unknown parameter, referred to as regression coefficient, y is known as explained variable, and x1,
x2......xmIt is the m general variances collected, i.e. explanatory variable.As m=1, as Linear Regression Model in One Unknown;m≥
It is then multiple linear regression model when 2, ε indicates random error.
For the N group monitoring data that we obtain, i-th group can be indicated are as follows: yi=xi1, xi2.....xim, wherein i=1,
2,3 ... n, then linear regression model (LRM) may be expressed as:
Since there are residual error e between actual value and estimated valuei=yi-(μ0+μ1xi1+…+μmxim+εi), utilize least square
The estimation technique improves the precision of estimated value so that the quadratic sum of residual error is minimum:
Σei 2=Σ (yi-(μ0+μ1xi1+…+μmxim+εi))2 (4)
Its necessary condition are as follows:
It is computed, formula (4) can convert are as follows:
To the deviation of estimated value estimated value and actual value be therefore,By left and right simultaneously multiplied by x ', can be obtainedAnd x ' e=0, therefore:
To the least-squares linear regression estimated value being calculatedNecessary inspection and evaluation are carried out, that is, is examined total
In body level, it whether there is linear relationship between dependent variable and independent variable, specifically include: the measurement of regression equation fitting degree,
Standard error estimate, the hypothesis testing of regression equation, the hypothesis testing of regression coefficient and multicollinearity differentiation etc..
Estimated value is calculated according to collected performance parameter sampleAnd verified by actual value, to obtain
Final regression equation:
Due toIt is the unbiased esti-mator of E (y), thus it is availableAs final predicted value.
2) periodic term prediction model
Periodic term has mechanical periodicity feature, and regular repetition, therefore, same phase are constantly done within certain time
On point can be in the variation up and down nearby of a specific value.To periodic term StIt is decomposed, the data point on same phase is just
Constitute sequence.
If periodic term sequence is { T (t), t=1,2 ... N }, the data point of cycle T=h, same phase constitutes same row,
Then former sequence can transform to following matrix:
Above-mentioned matrix (8) is split into the subsequence of h out of phase, by column vector so as to complete the weight in the period
Structure.Each subsequence is modeled, by changing rule of the simulation same phase in different cycles, this phase can be predicted
Data value of the position in next period, so that multi-step prediction is converted into Single-step Prediction.
3) random entry prediction model
In time series other than trend term and periodic term, it is also possible to there are certain equilibrium fluctuations, this time sequence
Column are usually stationary sequence.For stationary sequence, autoregressive moving-average model (Auto-Regressive and can be used
Moving Average Model, abbreviation arma modeling) Time Series Analysis Method model and predict.ARMA is as limited parameter
Model can determine model as long as determining limited parameter completely.The pre- of random entry subsequent time is calculated using arma modeling
Measured value mainly according to before its t moment p history value and q error amount be linearly calculated.ARMA (p, g) model specifically may be used
Description are as follows:
V (t+1) is the predicted value of subsequent time random entry in formula (9),With the undetermined coefficient that θ is not equal to zero, p and q
Respectively arma modeling Autoregressive and moving average order;V (t+1-i) is the random entry measured value before t moment, ε (t+
1-j) error term is predicted for the random entry before t moment.Therefore, model parameter p is most importantly determined in random entry prediction model
With the value of order q.
For the value of order q, there is nearly 95% numerical value to all fall within 2 after the auto-correlation coefficient of time series is in certain d rank
Within the scope of times standard deviation, and the wave process for decaying to small value of auto-correlation coefficient, then it is believed that model order d value thus.
The value of model parameter p is estimated using least square regression, obtained residual by acquiring N number of sample data
Difference are as follows:
Wherein, actual value when p=1,2 ..., N, V (t+1) are t+1, residual error e are the deviation of actual value and predicted value,
N is sample size.According to the least square regression that formula (4) trend term is predicted, its solution, which can be obtained, is
After the value of model parameter p and order q determines, then random entry can be predicted according to historical data.
It is bent that initial data is generated below with reference to trend prediction model algorithm flow chart (such as Fig. 1) based on present invention implementation
Line chart (such as Fig. 2), final obtain are coincide preferable prediction model prediction data figure (such as figure with truthful data compared with truthful data
3) this method is described:
1, network management performance choice of parameters
In optical transmission device, include on the influential performance parameter of the quality of service of communication network operating status and carrying
Background Block Error, errored block, errored seconds, the bit error rate, unavailable second, Severely Errored Second, continuius severely errored second, FEC correction word
Frame count, CRC check mistake, Pointer Justification Count, the Protection Switching Count PSC, operating temperature, input that section counts, FEC cannot be corrected
Optical power, Output optical power, laser bias current etc..Wherein, the bit error rate and optical power are to measure communication network signal transmission
The important indicator of quality.
2, network management performance data obtains
This section is by taking the optical power sending value of certain actual electric power saving communication network optical transmission device as an example, the period one
Characteristic value in year, collection period are 24 hours, obtain 365 groups of data altogether.Raw data plot figure is shown in Fig. 2.
3, the prediction model based on time series calculates
Trend prediction algorithm based on time series, using time series forecasting, steps are as follows:
1) trend term decomposes to obtain by sliding average algorithm, then carries out multinomial least square regression and obtain trend term
Prediction data, wherein order selects 8 times.
2) periodic term is obtained by multicycle superposed average, reuses the prediction that continuation method carries out subsequent cycle data.
3) random entry reuses ARMA (p, g) model (autoregressive moving-average model) and obtains random entry by obtaining
Prediction data.
4) by time series addition model, final predicted value is obtained.And prediction technique is verified by cross validation
Accuracy.
4, cross validation is carried out with true value
As seen from Figure 3, anticipation trend and the actual monitoring trend goodness of fit are higher, have preferable practicability.
Basic principles and main features and advantage of the invention have been shown and described above.The technical staff of the industry should
Understand, the present invention is not limited to the above embodiments, and the above embodiments and description only describe originals of the invention
Reason, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes and improvements
It all fall within the protetion scope of the claimed invention.The claimed scope of the invention is by appended claims and its equivalent circle
It is fixed.