CN102195814B - Method and device for forecasting and predicting by using relevant IT (Information Technology) operation and maintenance indexes - Google Patents

Abstract

The invention discloses a method and device for forecasting and predicting by using relevant IT (Information Technology) operation and maintenance indexes. The method comprises the following steps of: updating a data source, continuously receiving test data and updating a history database; performing necessary preprocessing on history data, calculating a covariance matrix, determining the relevance relation of indexes, decomposing and processing according to a characteristic value, and determining a covariance fluctuation range among the indexes; forecasting data: calculating a test covariance and forecasting according to the relevance and the fluctuation range; and predicating data: obtaining a linear equation set according to the relevance relation and the test covariance and obtaining a predicting result according to the equation set. Compared with the conventional intelligent forecasting and predicting method, the method has the advantages: the amount of calculation is reduced, and high forecasting accuracy and high predicting accuracy are obtained. The invention further discloses a device for forecasting and predicting IT operation and maintenance indexes.

Description

A kind of IT O﹠M indices prediction and Forecasting Methodology and device that utilizes correlation

Technical field

The present invention relates to IT operation management field, especially the monitoring between the index of IT O﹠M and management domain, be specifically related to a kind of index intelligent prediction of correlation and method of prediction utilized.

Background technology

The IT operation management, namely IT enterprises or department adopt relevant method, means, technology, system, flow process and document etc., the integrated management that IT running environment (comprising physical environment, hardware environment etc.), IT operation system and IT O﹠M personnel are carried out.Along with deepening continuously of building of IT and perfect, the operation maintenance of computer hardware and software system has obtained attention, because this is a new problem that produces along with the deep application of computer information technology, therefore how research carries out effective IT operation management, will have vast potential for future development and huge realistic meaning.

Briefly say, the organize content of IT O﹠M can manage and safeguard through being taken into index.Index, also namely describe the data of a certain characteristics of objects.The administration behaviour of IT O﹠M, can be abstracted into the change of data in essence.Therefore, the management of research IT O﹠M index highly significant.In the present invention, proposition is that a kind of correlation of utilizing is carried out the method for intelligent prediction and prediction to index.

Intelligent prediction and prediction, namely carry out the process of alarm or estimation by the mode of unartificial detection to certain desired value.Intelligentized example is a lot, is applied to the function of mobile phone or terminal hand-writing input method as the clustering algorithm with pattern recognition, can improve input efficiency; Some music software provides the function of automatic recommendation song for another example, by recording audience's historical record, predicts, this didactic mode can further meet audience's wish; 360 security guards are to the program updates of operating system with safeguard the function that automatic forecasting is provided for another example, can optimization system, improve system useful life.

Intelligentized theoretical system has developed to obtain comparative maturity, intelligent theoretical method and the means of application mainly comprise at present: (1) adaptation theory system, and this theory is a kind of feedback theory in essence, comprises the artificial neural net system, by the learning training sample, the predict future data; (2) area of pattern recognition, reach the purpose of identification by structure different mode system; (3) Optimum Theory system, this theory comprises supporting vector machine model, ant group algorithm, genetic algorithm, linearity and nonlinear restriction model reach the purpose of optimization aim data by modeling; (4) modern signal processing Domain Theory and method, signal processing method such as moving average adaptive regression model, and filtering method such as Wiener filtering, Kalman filter model, by modeling to the future time amount predict, level and smooth or estimate.

In the present invention, will directly not use above-described intelligent method, but utilize correlation.

Certainly exist correlation between some index of IT O﹠M.Detect as example take the WLAN index, the field intensity signal to noise ratio intensity of WLAN signal directly affects the network data bandwidth, even as connectedness such as the ping packet success rate of network, the Congestion Level SPCC of network may affect the WEB Authentication target, because when offered load is overweight, the WEB authentication access delay time may increase.In the application scenarios of reality, because of Cost Problems, some WLAN index should not be monitored constantly, as the field intensity signal to noise ratio, and some data can obtain constantly by the mode of software supervision, and between these two kinds of indexs or exist contact between more indexs, in this case, utilize the correlation between index just can overcome the problem that other Intelligent Plan is unpredictable or predictablity rate descends, because no matter whether data are known, correlation between index is constantly to exist, and only needs as adopting the method in the present invention just can reach the effect of prediction.In addition, correlation can also when some index unknown data dynamic range, forecast whether it exceeds standard.

The Mathematics Proof of correlation is as follows:

For two vectors

, covariance so between the two can be expressed as

（1）

Formed the matrix of the capable M row of M by the cross covariance between M index,

（2）

The definition coefficient correlation

, according to the character of coefficient correlation, auto-correlation coefficient equals that 0, two vector is uncorrelated, and the auto-correlation coefficient absolute value equals 1, and two SYSTEM OF LINEAR VECTOR that and if only if are relevant.Thus, we infer, the covariance absolute value is more more uncorrelated close to 0, two index, otherwise more relevant.

Summary of the invention

The invention provides a kind of IT O﹠M index intelligent forecasting and Forecasting Methodology of utilizing correlation, the feature of each step of the method is:

(1) upgrade Data Source, training data sample and test data sample data are provided, and wherein the training data of each index is multidimensional, and test sample book is one dimension,, along with passage of time, make training sample huge gradually after incorporating test sample book into historical data base.

(2) training, comprise data preliminary treatment and data and calculate two steps, and the burr data such as minimax can be eliminated in the training sample source after the data preliminary treatment, reach smooth effect, thereby for next step, provide accurately reasonably Data Source; During data process data calculation procedure after pretreatment, obtain a covariance matrix according to formula (1), (2), then calculate the covariance fluctuation range.

Preferably, at first, matrix (2) is done Eigenvalues Decomposition obtain

（3）

,

Be respectively characteristic vector and characteristic value diagonal matrix, then, keep the larger characteristic value of absolute value, reject little order and equal zero, thereby obtain

, so,

（4）

Inevitable is also a symmetrical matrix, and differs from

, consider the element of triangular portions on it, define fluctuation range and be: a boundary of fluctuation range

, another boundary is so

(5)

(3) test, comprise data forecast and two steps of data prediction.In data forecast step,

Preferably, at first, from data source, obtain test sample book, fluctuation range and i and j index average of any two indexs that obtain according to training module

, the covariance that defines between any two test sample book data is expressed as,

（6）

Can judge Whether drop on

Fluctuation range in, thereby forecast.

Preferably,, if known a certain index but can't forecast whether it exceeds standard, forecast that thought is: find draw in training module with the maximally related several indexs of this index,, if one of them index can be forecast sequentially, stop forecast.

Under the prerequisite that can't detect achievement data, can predict index.

Preferably, according to formula (6), the algorithm of Accurate Prediction is: first find the maximally related index j with index i to be measured, then find the maximally related index k with j, can think

, the equation left side is unknown test covariance, the right is known training covariance.Thereby three systems of linear equations of simultaneous, separate namely obtain to predict the outcome also namely separated.Be also that solving equations obtains X

（7）

The present invention also provides a kind of intelligent forecasting and prediction unit that utilizes correlation simultaneously, comprises,

The data source module,, with the initialization data of existing historical data as training module, select as far as possible large.Simultaneously,, for the test data of constantly updating, incorporate it into tranining database after often testing one group of data, guarantee upgrading in time of database.

Preferably, when data volume reaches certain scale, carry out the packet training, to improve test accuracy.Referring to key diagram 1.

Training module, comprise data pretreatment unit and data computing unit,

The data pretreatment unit,

Preferably, eliminate the burr purpose in order to reach,, to each index, under initial situation, first remove obviously extreme several sample values and keep remaining sample, calculate as several extremely large arithmetic mean M and several extremely little value arithmetic mean m,, when at every turn more during new data,, if find that data drop on outside M or m, be regarded as burr and reject, the data group of rejecting simultaneously forms new manifold, upgrades M and m.Go in such a manner, make data reach level and smooth effect as far as possible.Shown in Figure 5 referring to explanation.

The data computing unit,

Preferably, because the data preprocessing part is eliminated the burr processing to each index, may make between two achievement data vectors dimension different, the mode that solves is, for burr of an every elimination of index, when shortage of data, replace the error while to reduce, calculating covariance matrix with the arithmetic mean of all data acquisition systems of front;

Preferably, the rule of rejecting less characteristic value is,, with the addition that takes absolute value of all characteristic values, then calculates the ratio of each characteristic value, if this characteristic value ratio less than as 0.05, claim that the characteristic value contribution margin is too small, also it equals zero even it can be considered to reject.Reject manyly, the fluctuation range of calculating is larger.This execution mode can be shown in Figure 6 referring to explanation.

Test module, comprise data forecast unit and data prediction unit,

Data forecast unit, comprise and differentiating and forecast,

Differentiate, in reality, in a single day some index is measured just has with reference to scope, therefore need not forecast, and for the other index, measures not with reference to scope, and whether therefore at first distinguish index needs to forecast;

Preferably, the algorithm principle of forecast module is to see first whether with maximally related that index of index x to be measured be index known and in known dynamic range, if not continue search, meet the demands until search out front m, the m maximum can reach all known dynamic range index numbers.First is made as i, and index i and x are calculated covariance conv (x, i), if less than fluctuation range, forecasting index x does not exceed standard; If otherwise greater than fluctuation range, calculate again index j with index x correlations, if conv (j, i), less than fluctuation range, forecasts that x exceeds standard, otherwise, claim the i prediction to lose efficacy, replace i with j, repeat the flow process of i.So repeatedly,, until before all, m index all predicted inefficacy, forecast that x does not exceed standard.

This unit specifically can be shown in Figure 7 referring to explanation.

The data prediction unit, be used for prediction some can't direct-detection data, be divided into and differentiate and prediction, preferably, according to mentioning the thought of solving an equation in method, carry out.Referring to key diagram 8.

The flow chart of whole device is as shown in Figure 4 in illustrating.

A kind of IT O﹠M index intelligent forecasting and Forecasting Methodology and device that utilizes correlation provided by the invention, its intelligent being embodied in: in the time of can't judging when the given data source whether it exceeds standard, usage data test cell, alarm in the IT of reality O﹠M system; In the time due to chance failure or additive method, can't directly to certain index, detecting, utilize all the other associated desired values and data prediction unit, can predict more exactly this index.

A kind of IT O﹠M index intelligent forecasting and Forecasting Methodology and device that utilizes correlation provided by the invention, its advantage and characteristics are: with traditional intelligent prediction or Forecasting Methodology, compare, all need two steps of training and testing, but amount of calculation is much smaller, and can reaches higher accuracy.

Description of drawings

Examples of the present invention will be described by way of reference to the accompanying drawings, wherein

Fig. 1 is every group of number of station work and certain index success rate graph of a relation of prediction.

Fig. 2 is the magnitude relationship figure of a certain test index warning probability and this index.

Fig. 3 is the variation relation figure of the predicted value deviation ratio of a certain test index with the index size.

Fig. 4 is the flow chart of device.

Fig. 5 is the flow chart of the data pretreatment unit of training module.

Fig. 6 is the flow chart of data computing unit in training module.

Fig. 7 is the flow chart of data forecast unit in test module.

Fig. 8 is the flow chart of data prediction unit in test module.

Fig. 9 is whole method and apparatus system principle schematic diagram.

Embodiment

, for making the inventive method and device can reach result and the function of expectation,, simultaneously for more clear and intuitive expression method of the present invention, will adopt the simulation result figure of MATLAB describe and show.

In specific embodiment 1, with reference to key diagram 1,

Suppose under real scene, receive altogether 20 achievement data sources, the statistical history data, suppose that the initial sampled data of each index is fixed as 1000, and establishing index training data to be measured source is that average is 10, and variance is 0.1 just too distributed data.Consideration divides into groups to enter the processing of training module to it, in theory, for guarantee that fluctuation range calculates accurately, every group of number is unsuitable very few, simultaneously for smoothing processing, the group number should not very little, therefore, have a compromise.This routine purpose is how checking fixedly the time, distributes these data can reach good performance when data source.For embodiment 2 does foundation.

Shown in Figure 1 by explanation, under the testing data known cases, set two kinds of situations:

Index test data to be measured equal 10, and in scope, presentation of results, be divided into when 1000 data under the scope of every group 100 ~ 500, and predicated error is lower than 0.1; Test data equals 14 outside scope, and presentation of results when 1000 data are divided into every group 100 ~ 500, can reach better prediction effect relatively, and predicated error is minimum in 0.4 left and right.

By embodiment 1, obtain the allocation proportion of 1000 number grouping numbers and group number, can elect 100 every group as, totally 10 groups, as the foundation of next embodiment.

Simultaneously, this example also illustrated on duty exceeded scope after, its predicted value is very inaccurate, this explanation several indexs relevant to this index have all exceeded scope, because the satisfied condition of predicting, so this situation does not meet application category of the present invention.

In specific embodiment 2, see also key diagram 2.

Suppose under real scene, several 10 of index, training data adds up to 10000, it is divided into 100 groups, every group of 100 data, the training data source is the random number between 0 ~ 1, preset desired value to be measured often increases progressively 0.5 until approach 20 from 0, puts 1 for reporting to the police (exceeding standard), and 0 does not report to the police.In theory, when data are got over away from this scope of 0 ~ 1, report to the police and should be 1, otherwise be 0.The algorithm robustness that provides due to method exists, so, after smoothing processing, by the warning probability, reflect the forecast performance.

Shown in Figure 2 by explanation, gradually away from 1 the time, the warning probability rises gradually, until approach 1 when initialize data (wait forecasting test data).In reality, the mode of solution is, sets up a threshold values, reports to the police higher than threshold values when certain test data obtains the warning probability, otherwise do not report to the police.

This embodiment has verified the validity of inventive method data forecasts, and a solution is provided.

In specific embodiment 3, see also key diagram 3.

Suppose under real scene, the index number is 20, and there are 1000 data in every group of index training data source, and achievement data to be measured source is take 10 as average, and 0.1 is the random number of variance, and preset index test data to be measured are incremented to 15, calculating prediction deviation rate from 5 with 0.5.

Shown in Figure 3 by explanation, when presetting range during in 10 scope, the I of predicated error is lower than 0.1, otherwise predicated error is increasing.This key diagram, the same manner as in Example 1, illustrated that the Forecasting Methodology that the present invention provides has higher precision.

Claims (1)

1. IT O﹠M indices prediction and a Forecasting Methodology of utilizing correlation, is characterized in that, described method comprises three key steps: upgrade Data Source, training and testing;

Upgrade Data Source, specifically comprise with initialized historical data as training data, constantly incorporate test data simultaneously into, the renewal historical data base;

Training, specifically comprise data preliminary treatment and data calculation procedure, the data preliminary treatment, and the input training data, eliminate the burr data and carry out smoothing processing; Described data calculation procedure, calculate covariance matrix, determines the correlative relationship between index, it carried out Eigenvalues Decomposition process definite fluctuation range;

Covariance matrix utilizes the covariance formula to obtain, and the computing formula of covariance is: if x _i=[x ₁..., x _N], x _j=[x' ₁..., x' _N] expression i and j achievement data, both covariances can be expressed as so $\mathrm{conv}(x_i,x_j)=\frac{x_i\·{x^{\′}}_j}{N}-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\frac{x_i}{N}*\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\frac{x_j^{\′}}{N},$ This value is more close to zero, and two indexs are more uncorrelated, otherwise more relevant, thereby determines index related relation; After calculating covariance matrix, it is carried out Eigenvalues Decomposition, with the larger reservation of absolute value in all characteristic values, rejecting is close to zero part, again revert to new covariance matrix, establishing the covariance size of i and j index in the new covariance matrix that recovers after the past, characteristic value was processed is conv'(x again _i, x _j), establishing it and be circle of fluctuation range, another boundary is conv''(x so _i, x _j)=2*conv (x _i, x _j)-conv'(x _i, x _j), be conv'(x thereby obtain any two indexes covariance fluctuation range _i, x _j)～conv''(x _i, x _j);

Test, specifically comprise data forecast and data prediction step, and the data forecast,, by correlative relationship and the fluctuation range that obtains after training, utilize index of correlation to forecast whether it exceeds standard; Described data prediction, solve and predict the outcome by correlative relationship and index of correlation value Simultaneous Equations;

Condition and the implication of described data forecast are: the data of index to be measured have recorded and under the situation of non-burr data, the standard of there is no goes to define its excursion, therefore utilize correlation to forecast indirectly whether it exceeds standard; The algorithm steps of data forecasts is: known a certain test index value X, first find maximally related other several indexs of desired value X that obtain with this test that draw in training module, and needing only some indexs according to sequencing can forecast, stops forecast;

The condition of described data prediction is: these data can't detect and obtain by direct mode due to fault or other reasons, and all the other associated indexs can record and all not exceed dynamic range; The method of prediction is: first find the maximally related index j with index i to be measured, then find the maximally related index k with j, can think conv _ij-conv _ik=CONV _ij-CONV _ik, the equation left side is unknown test covariance, the right is known training covariance, even also the test covariance equates with the training covariance, as equation 1; Then utilize covariance to obtain i and j, and the covariance accounting equation of i and k

With

Thereby obtain containing three equations of three parameters, that separates namely obtains and predicts the outcome.

Images