CN106384128A - Method for mining time series data state correlation - Google Patents

Abstract

The invention relates to a method for mining time series data state correlation, and the method comprises the steps: preprocessing including outlier removing, uniformly-spaced interpolation, normalization operation and the like is performed on time series data variables; state mining is performed on single variable, a dynamic partitioning and clustering method is adopted to cluster integrated eigenvectors of all windows of the single variable, windows from different clusters represent different states, all clusters are ranked according to sizes, each window is represented by a character corresponding to the belonged cluster, then original numeric data is transformed to character string form, that is, a state character string of each variable can be acquired; the state character strings of all variables are aligned, and a multivariable state matrix is obtained; an Apriori algorithm is utilized to mine correlation rules between different variable states and bring a formal representation and correlation intension thereof out; the correlation rules are reduced to remove redundant information at last; according to the invention, the method has an ability of noise disturbance resistance and is suitable for carefully analyzing state value correlation of a small parameter set, and a state value mapping relation can be mined.

Description

A kind of method excavating time series data state relation

Technical field

The invention belongs to Intelligent Information Processing and field of computer technology are and in particular to a kind of time series data state of excavating is closed The method of connection.

Background technology

In large-scale complicated system, between variable states, comprise certain incidence relation.This incidence relation is subject to system The effect of inherent laws, just has certain embodiment on abnormal data.Relevance can show as on space-time cooccurrence relation, because Fruit relation, tendency relation, correlation etc..When system mode changes, the respective change of different variables will be caused.System Under normal condition with abnormality, moving law is different, is reflected in the version difference showing as variable on abnormal data. By analyzing the Changing Pattern of multiple variable abnormal datas, excavate the relevance between different variable states, for total clone System moving law, finds that incipient fault knowledge has important function.

Content of the invention

It is an object of the invention to provide a kind of method excavating time series data state relation, the method integrated use spy Levy extractive technique, clustering learning theory to excavate the state of single variable, then utilize Apriori algorithm to excavate different variables State relation rule simultaneously provides formalization representation and strength of association, finally correlation rule is carried out with yojan and eliminates redundancy；This Ambiguity or the uncertainty of variable-value is considered in invention, has noise resistance interference performance, is suitable for thin to small parameter set Cause ground to analyze its state value relevance, excavate state value mapping relations.

For reaching above-mentioned purpose, the technical solution used in the present invention is：

A kind of method excavating time series data state relation, the system realizing the method includes data preprocessing module, spy Levy extraction module, dynamic partition clustering module, multivariable state matrix generation module, Apriori state relation excavate module and Correlation rule yojan module, it comprises the concrete steps that：

1) first, data preprocessing module carries out elimination of burst noise, at equal intervals interpolation, normalization operation to original temporal data, Obtain valid data form；

2) secondly, the valid data of time series variable are divided into the window of equal length by characteristic extracting module, to every Individual window data extracts feature, including Fourier's feature, statistical nature, wavelet character constitutive characteristic vector；

3) and then, dynamic partition clustering module enters Mobile state partition clustering to the characteristic vector of all windows of single variable, The cluster obtaining will be clustered in magnitude order, maximum cluster is represented with character ' a ', secondary big cluster is represented with ' b ', the like, Then it is considered as noise less than the cluster of given threshold value 2, with '？' represent, by each window with its corresponding character representation of place cluster, so Raw value type data is converted to character string forms, i.e. the status word string of this variable；

4) multivariable state matrix generation module, by the status word string of all variables according to time unifying, forms state square Battle array；

5) Apriori state relation excavate module Apriori algorithm multivariable state matrix is carried out frequent item set and Association rule mining；

6) last, correlation rule yojan module carries out yojan to the correlation rule detecting, and eliminates redundancy, obtains Whole multivariable state relation rule.

The step of described data preprocessing module elimination of burst noise includes：Calculate average and the standard deviation of each window, judge The standard deviation of the difference of each data point and watch window average that it the is located whether watch window more than 5 times, if being more than, this number Strong point is unruly-value rejecting；Interpolation at equal intervals is carried out to the time series after elimination of burst noise, if being spaced apart △ t, initial time is T, then Time set after interpolation should be { T+n* △ t n=0,1,2,3 ... } at equal intervals, and T+i* △ t is corresponding to be worth for original sequence In row from this moment nearest less than the value corresponding to T+i* △ t, that is, in original series first be more than T+i* △ t The observation corresponding to the previous moment；Linear normalization is carried out to the data after interpolation operation at equal intervals, scans one first All over time series, obtain maximum (max) and the minimum of a value (min) of observation, each observation station is calculated according to equation below and returns Numerical value after one change, original time series span is transformed on [0,1] interval；

$x_i=\frac{x_i-min}{\mathrm{\Δ}}$

Wherein, x_iRepresent i-th observation station numerical value；△=max-min.

Characteristic extracting module：First, with the window setting, univariate data is cut；Secondly, in each window Data carry out feature extraction, including statistical nature, Fourier's feature, wavelet character；V1=[average, variance], v1 represents system Meter feature, wherein average reflects the average level of data in a window, and variance then describes the fluctuation journey of data in window Degree；V2=[Fourier coefficient 1, Fourier frequency 1, Fourier coefficient 2, Fourier frequency 2], v2 represents frequency domain character, passes through Fourier transformation obtains a series of Fourier coefficient, and Fourier coefficient is ranked up from big to small according to absolute value, chooses The maximum Fourier coefficient of the first two and its corresponding frequency；V3=[wavelet coefficient detail coefficients 1 ... wavelet details coefficient N], v3 represents time and frequency domain characteristics, carries out wavelet transform to each window, obtains n wavelet details coefficient, by this three aspect Characteristic synthetic gets up, and constitutes window feature vector v, v=[v1, v2, v3].

Described dynamic partition clustering module adopt dynamic partition clustering method to the feature of all windows of single variable to Amount carries out clustering its process as follows：

1) the independent cluster of first window, the cluster heart is the multi-feature vector of this window；

2) the initial division process of cluster, calculates similar between the 2nd window and the 1st cluster cluster heart according to equation below Degree：

$cos(v_{ik},v_{jk})=\frac{v_{ik}\·v_{jk}}{\left|v_{ik}\right|\×\left|v_{jk}\right|}$

In formula：cos(v_ik, v_jk) represent window i v_kThe v of (k=1,2,3) vector sum window j_kCosine between vector is similar Degree；cos(v_ik, v_jk) ∈ [- 1 ,+1], using equation below, calculate the distance between two windows；

$d(v_{ik},v_{jk})=1-\frac{cos(v_{ik},v_{jk})+1}{2}$

$dist(v_i,v_j)=\frac{\underset{k=1}{\overset{3}{\mathrm{\Σ}}}dist(v_{ik},v_{jk})}{3}$

In formula：dist(v_i, v_j) represent the distance between window i and j, dist (v_i, v_j) ∈ [0,1]；

If dist<D, d=0.2), then No. 2 windows are incorporated to first cluster and the cluster heart changed immediately according to equation below：

${\mathrm{cv}}_k=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{v}_{ik}}{n},(k=1,2,3)$

In formula：cv_kRepresent cluster heart c k-th characteristic vector, it be equal to cluster in all k-th characteristic vectors of window equal Value, the cluster heart c of a cluster is exactly cv₁, cv₂, cv₃Combination, that is, in cluster all window multi-feature vectors mean value；If Dist >=d, the independent cluster of No. 2 windows, process remaining window successively：Calculate i-th window and all cluster cluster hearts having produced The distance between, pick out the cluster c nearest with it apart from dist, if dist<D, i window is incorporated to cluster c, otherwise individually becomes Cluster；

3) the adjustment process of cluster：(i=1,2 ... m), calculates the cluster heart distance of itself and all clusters to take out No. i-th window Dist, picks out the dist and its corresponding cluster c of minimum, if dist≤d, and window i is in cluster c, then by window i from original Cluster move on to cluster c；If dist≤d, and window i is in cluster c, then window i is not operated；If dist>D, then by window i from former Remove in the cluster coming, independent cluster；Repeat said process, until having processed all windows, calculate the cluster heart of all clusters.If existing One cluster heart there occurs change, then repeat the adjustment process of cluster, i.e. step 3) until the cluster heart of all clusters no longer changes；If all The cluster heart is all constant, then execution step 4)；

4) merging process of cluster：Calculate the cluster heart distance of any two cluster, select closest two cluster c_i, c_j, and its Corresponding apart from dist, if dist≤α, α=0.3), then merge cluster c_i, c_j, and calculate the cluster heart of new cluster after merging, repeat to close And process 4), if dist>α, represents that there are not two close enough clusters can merge, then exit merging process, clustering algorithm Terminate, the window feature in same cluster in cluster result approximate it is considered to be a kind of state, the window of different clusters represents difference State；By all clusters in magnitude order, maximum cluster is represented with character ' a ', and secondary big cluster is represented with ' b ', the like, Then it is considered as noise less than the cluster of given threshold value, with '？' represent, by each window with its corresponding character representation of place cluster, so Raw value type data is converted to character string forms, that is, obtain the status word string of each variable.

Described multivariable state matrix generation module be by the status word string of all variables according to time unifying it is assumed that There is n variable, the initial observation time of each variable is identical with deadline, then their window number is necessarily identical, shape State string length is also identical it is assumed that status word string length is m, then generate the multivariable state matrix of n*m.

It is to enter line frequency using Apriori algorithm to multivariable state matrix that described Apriori state relation excavates module Numerous item collection and association rule mining：Apriori algorithm Mining Frequent Itemsets Based flow process is as follows：First, by scanning transaction journal, look for Go out all of frequent 1 item collection, this set is denoted as L₁, then utilize L₁Find frequent 2 item collections L₂, so on, until looking for again To any frequent k item collection；Iteration is divided into two steps every time：First, produce candidate by connecting step and beta pruning step；The Two, calculate the support of each candidate item, be considered as frequent episode more than the item of minimum support threshold value 0.001, in frequent episode Mining Association Rules on the basis of collection, specific as follows：Firstly, for each frequent item set L, produce all nonvoid subsets of L；Its Secondary, for each nonvoid subset s of L, produce a candidate rule " s → (L-s) ", after wherein (L-s) represents removing s in L Remaining content；If the confidence level of this candidate rule last is more than given threshold value 0.5, exports this rule, otherwise abandon this Rule, the confidence level of rule is calculated as follows：

$Cf(L,s)=\frac{Sp\left(L\right)}{Sp\left(s\right)}$

Wherein, (L, s) represents the confidence level of regular " s → (L-s) " to Cf, and Sp (L) represents the support of L, and Sp (s) represents s Support.

Described correlation rule yojan module merges to the redundancy rule producing or deletes, and reduction steps are：

1) for the correlation rule obtaining, it is ranked up from big to small according to confidence level；

2) for each K rank frequent episode (K>1), only retain the maximum correlation rule of confidence level；

3) if the former piece of two correlation rules is identical, consequent is compared, if there is inclusion relation in consequent, On the premise of confidence level difference very little, delete the rule belonging to consequent being comprised；

4) if the consequent of two correlation rules is identical, former piece is compared, if there is inclusion relation in former piece, On the premise of confidence level difference very little, delete the more rules of former piece, retain the fewer rule of former piece；

5) in order to ensure the uniformity of knowledge, it is to avoid circular reasoning occurs, need to detect with the presence or absence of ring in correlation rule, Detection to ring is realized by directed acyclic graph, represents former piece with a node, and a node represents consequent, and the two uses directed edge Connect, detect the correlation rule according to confidence level descending one by one.

With respect to prior art, the present invention first, detects the state of each variable, using dynamic partition clustering method to institute Window characteristic vector is had to be clustered, the window feature in same cluster is approximate, represents same state, the window in different clusters Represent different states.By the status word string of all variables according to time unifying, obtain multivariable state matrix.Using Apriori algorithm excavates the frequent cooccurrence relation between different variable states values, thus obtain multiple variables difference values it Between association, and provide Formal Representation and its strength of association.Finally remove redundancy with the correlation rule producing is carried out with yojan Information.The present invention take into account ambiguity or the uncertainty of variable-value, has noise resistance interference performance, is suitable for little ginseng Manifold is closed and is meticulously analyzed its state value relevance, excavates state value mapping relations.

Brief description

Fig. 1 is the module frame figure of present system.

Fig. 2 is the present invention dynamic partition clustering block flow diagram.

Table 1 is the status word string of example sequential variable of the present invention.

Table 2 is the state relation rule digging result of example sequential variable of the present invention.

Fig. 3 is few examples sequential variable states correlation rule schematic diagram of the present invention.

Fig. 4 is few examples sequential variable states correlation rule schematic diagram of the present invention.

Specific embodiment

Below in conjunction with the accompanying drawings the present invention is described in further detail.

Referring to Fig. 1, the system realizing the present invention includes data preprocessing module 1-1, characteristic extracting module 1-2, dynamically draws Segregation generic module 1-3, multivariable state matrix generation module 1-4, Apriori state relation excavate module 1-5 and correlation rule Yojan module 1-6；The comprising the concrete steps that of the inventive method：

1) first, data preprocessing module 1-1 carries out elimination of burst noise, at equal intervals interpolation, normalization behaviour to original temporal data Make, obtain valid data form；

The step of elimination of burst noise includes：Calculate average and the standard deviation of each window, judge each data point and sight that it is located Examine the standard deviation of the difference of the window average whether watch window more than 5 times, if being more than, this data point is unruly-value rejecting；To going Time series after outlier carries out interpolation at equal intervals, if being spaced apart △ t, initial time is T, then the time collection after interpolation at equal intervals Conjunction should be { T+n* △ t n=0,1,2,3 ... }, T+i* △ t is corresponding be worth in original series from this moment nearest little In the value corresponding to T+i* △ t, that is, corresponding to first previous moment more than T+i* △ t in original series Observation；Linear normalization is carried out to the data after interpolation operation at equal intervals, scans a time series first, obtain observation Maximum (max) and minimum of a value (min), according to equation below calculate each observation station normalization after numerical value, when will be original Between sequence span be transformed on [0,1] interval；

$x_i=\frac{x_i-min}{\mathrm{\&Delta;}}$

Wherein, x_iRepresent i-th observation station numerical value；△=max-min.

2) secondly, the valid data of time series variable are divided into the window of equal length by characteristic extracting module 1-2, right Each window data extracts feature, including Fourier's feature, statistical nature, wavelet character constitutive characteristic vector；

First, with the window setting, univariate data is cut；Secondly, feature is carried out to the data in each window Extract, including statistical nature, Fourier's feature, wavelet character；V1=[average, variance], v1 represents statistical nature, wherein average Reflect the average level of data in a window, variance then describes the degree of fluctuation of data in window；V2=[Fourier system Number 1, Fourier frequency 1, Fourier coefficient 2, Fourier frequency 2], v2 represents frequency domain character, obtains one by Fourier transformation The Fourier coefficient of series, is ranked up according to absolute value from big to small to Fourier coefficient, chooses in maximum Fu of the first two Leaf system number and its corresponding frequency；V3=[wavelet coefficient detail coefficients 1 ... wavelet details coefficient n], v3 represent that time-frequency domain is special Levy, wavelet transform is carried out to each window, obtain n wavelet details coefficient, this tripartite's region feature is integrated, constitute Window feature vector v, v=[v1, v2, v3].

3) and then, dynamic partition clustering module 1-3 the characteristic vector of all windows of single variable is entered Mobile state divide poly- Class, will cluster the cluster obtaining in magnitude order, and maximum cluster is represented with character ' a ', and secondary big cluster is represented with ' b ', class successively Push away, be then considered as noise less than the cluster of given threshold value 2, with '？' represent, by each window with its corresponding character representation of place cluster, So raw value type data is converted to character string forms, i.e. the status word string of this variable；

Step 2-1 is first carried out, the independent cluster of first window, the cluster heart is the multi-feature vector of this window.Execute step Rapid 2-2, takes off a data.Execution step 2-3, calculates the distance of this data and all cluster hearts.Execution step 2-4, picks out The cluster c nearest with it, the distance between they are denoted as dist：

$cos(v_{ik},v_{jk})=\frac{v_{ik}\&CenterDot;v_{jk}}{\left|v_{ik}\right|\&times;\left|v_{jk}\right|}$

In formula：cos(v_ik, v_jk) represent window i v_kThe v of (k=1,2,3) vector sum window j_kCosine between vector is similar Degree.cos(v_ik, v_jk) ∈ [- 1 ,+1], using equation below, calculate the distance between two windows.

$d(v_{ik},v_{jk})=1-\frac{cos(v_{ik},v_{jk})+1}{2}$

$dist(v_i,v_j)=\frac{\underset{k=1}{\overset{3}{\mathrm{\&Sigma;}}}dist(v_{ik},v_{jk})}{3}$

In formula：dist(v_i, v_j) represent the distance between window i and j, dist (v_i, v_j) ∈ [0,1].

Execution step 2-5, judges dist and given threshold value d, d=0.2) relation, if dist≤d, execution step 2-6, This data is incorporated in cluster c and changes the cluster heart immediately；

${\mathrm{cv}}_k=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{v}_{ik}}{n},(k=1,2,3)$

In formula：cv_kRepresent cluster heart c k-th characteristic vector, it be equal to cluster in all k-th characteristic vectors of window equal Value, the cluster heart c of a cluster is exactly cv₁, cv₂, cv₃Combination, that is, in cluster all window multi-feature vectors mean value；

If dist>D, execution step 2-7, this data sheet alone becomes cluster and as the cluster heart.Execution step 2-8, judges whether All data are processed.If not processed, execution step 2-2, take off a data；Otherwise execution step 2-9, takes first Data.Execution step 2-10, the cluster heart calculating this data with all clusters is apart from dist.Execution step 2-11, picks out with it Near cluster c, the distance between they are denoted as dist.Execution step 2-12, judges the relation of dist and given threshold value d, if dist ≤ d, execution step 2-13, judge this data or not in cluster c, if this data does not exist in cluster c, execution step 2-14, should Data moves in cluster c；Otherwise execution step 2-15, does not operate to this data.If dist>D, execution step 2-16, this number According to independent cluster.Execution step 2-17, judges whether to have processed all data.If not processed, execution step 2-18, take off One data；Otherwise execution step 2-19, calculates the cluster heart of all clusters.Execution step 2-20, judges whether the cluster heart changes, that is, Whether cluster result changes.If changing, execution step 2-9；Otherwise execution step 2-21, selects closest Two clusters.Execution step 2-22, judges the distance between this two cluster hearts dist and given threshold value α, α=0.3) size.If dist<α, the cluster heart of new cluster, then execution step 2-21 after merging this two clusters and calculating merging；If dist >=α, exit Cluster process.Window feature in same cluster in cluster result approximate it is considered to be a kind of state, the window of different clusters represents Different states；By all clusters in magnitude order, maximum cluster is represented with character ' a ', and secondary big cluster is represented with ' b ', successively Analogize, be then considered as noise less than the cluster of given threshold value 2, with '？' represent, by each window with its corresponding character list of place cluster Show, such raw value type data is converted to character string forms, that is, obtain the status word string of each variable.

With reference to table 1, it is the status word string mining result to all example variables for the dynamic partition clustering module.To each Variable, the window feature in same cluster in cluster result approximate it is considered to be a kind of state, the window of different clusters represents difference State；By all clusters in magnitude order, maximum cluster is represented with character ' a ', and secondary big cluster is represented with ' b ', the like, Then it is considered as noise less than the cluster of given threshold value, with '？' represent, by each window with its corresponding character representation of place cluster, so Raw value type data is converted to character string forms, that is, obtain the status word string of each variable.

Table 1

4) multivariable state matrix generation module 1-4, by the status word string of all variables according to time unifying, forms shape State matrix；

Multivariable state matrix generation module 1-4 is it is assumed that there being n by the status word string of all variables according to time unifying Individual variable, the initial observation time of each variable is identical with deadline, then their window number is necessarily identical, status word Symbol string length is also identical it is assumed that status word string length is m, then generate the multivariable state matrix of n*m.

5) Apriori state relation excavation module 1-5 carries out frequent item set with Apriori algorithm to multivariable state matrix And association rule mining；

It is to enter line frequency using Apriori algorithm to multivariable state matrix that described Apriori state relation excavates module Numerous item collection and association rule mining：Apriori algorithm Mining Frequent Itemsets Based flow process is as follows：First, by scanning transaction journal, look for Go out all of frequent 1 item collection, this set is denoted as L₁, then utilize L₁Find frequent 2 item collections L₂, so on, until looking for again To any frequent k item collection；Iteration is divided into two steps every time：First, produce candidate by connecting step and beta pruning step；The Two, calculate the support of each candidate item, be considered as frequent episode more than the item of minimum support threshold value 0.001, in frequent episode Mining Association Rules on the basis of collection, specific as follows：Firstly, for each frequent item set L, produce all nonvoid subsets of L；Its Secondary, for each nonvoid subset s of L, produce a candidate rule " s → (L-s) ", after wherein (L-s) represents removing s in L Remaining content；If the confidence level of this candidate rule last is more than given threshold value 0.5, exports this rule, otherwise abandon this Rule, the confidence level of rule is calculated as follows：

$Cf(L,s)=\frac{Sp\left(L\right)}{Sp\left(s\right)}$

Wherein, (L, s) represents the confidence level of regular " s → (L-s) " to Cf, and Sp (L) represents the support of L, and Sp (s) represents s Support.

6) last, correlation rule yojan module 1-6 carries out yojan to the correlation rule detecting, and eliminates redundancy, obtains To final multivariable state relation rule.

1) for the correlation rule obtaining, it is ranked up from big to small according to confidence level；

2) for each K rank frequent episode (K>1), only retain the maximum correlation rule of confidence level；For example：Second order frequent episode (A, B), if producing correlation rule A → B and B → A, only retains of confidence level maximum；

3) if the former piece of two correlation rules is identical, consequent is compared, if there is inclusion relation in consequent, On the premise of confidence level difference very little, delete the rule belonging to consequent being comprised；For example：A → B and A → B, C, if | Cf (A → B)-Cf (A → B, C) |<δ, then delete A → B, retains A → B, C；

4) if the consequent of two correlation rules is identical, former piece is compared, if there is inclusion relation in former piece, On the premise of confidence level difference very little, delete the more rules of former piece, retain the fewer rule of former piece；For example：A → B and A, C → B, if | Cf (A → B)-Cf (A, C → B) |<δ, then delete A, C → B, retains A → B；

5) in order to ensure the uniformity of knowledge, it is to avoid circular reasoning occurs, need to detect with the presence or absence of ring in correlation rule, Detection to ring is realized by directed acyclic graph, represents former piece with a node, and a node represents consequent, and the two uses directed edge Connect, detect the correlation rule according to confidence level descending one by one.It is assumed that currently considering A → B, look first at directed acyclic in figure Whether A can be reached from B, that is, detect whether comprise A in all descendant nodes of B, if the descendant node of B comprises A, that is, oriented Figure comprises B → A, then delete A → B.If not comprising A in the descendant node of B, then interpolate A → B in digraph.

With reference to table 2, it is the result after variable states correlation rule yojan.

Table 2

With reference to Fig. 3, it is the schematic diagram of correlation rule " P002=b → P004=b ", and in figure dotted line (upper part) represents change ' b ' state of amount P002, wherein some blank, blank is not without data, but this window corresponds to other states, no It is ' b ' state.Solid line (lower part) is ' b ' state of variable P004.The confidence level of this rule is 47/50=0.940, represents The number of times that variable P002 and P004 goes out present condition ' b ' simultaneously in 146 records is 47, and ' b ' state individually in variable P002 Number of times be 50.

With reference to Fig. 4, it is the schematic diagram of correlation rule " P073=b → P075=b ", and wherein dotted line (top half) represents ' b ' state of variable P075, solid line (the latter half) represents ' b ' state of variable P073.' b ' state of P073 can be seen not Identical to the greatest extent, spike in some tops, and some tops are parallel lines.In fact, when excavating variable states, to this variable All window feature vectors are clustered, the window feature vector approximation (might not be identical) in same cluster, same Window in individual cluster represents a kind of state, so the feature of data, form are approximate in the corresponding window of state, but not complete Exactly the same.The confidence level of this rule is 44/44=1.0, represents that variable P073 and P075 goes out present situation simultaneously in 146 records The number of times of state ' b ' is 44, and the number of times that ' b ' state individually in variable P073 is also 44.' b ' state in this explanatory variable P073 When be always accompanied by ' b ' state of P075.

Claims (7)

1. a kind of method excavating time series data state relation is it is characterised in that the system of realizing the method includes data and locates in advance Reason module (1-1), characteristic extracting module (1-2), dynamic partition clustering module (1-3), multivariable state matrix generation module (1- 4), Apriori state relation excavates module (1-5) and correlation rule yojan module (1-6), and it comprises the concrete steps that：

1) first, data preprocessing module (1-1) carries out elimination of burst noise, at equal intervals interpolation, normalization operation to original temporal data, Obtain valid data form；

2) secondly, the valid data of time series variable are divided into the window of equal length by characteristic extracting module (1-2), to every Individual window data extracts feature, including Fourier's feature, statistical nature, wavelet character constitutive characteristic vector；

3) and then, dynamic partition clustering module (1-3) enters Mobile state partition clustering to the characteristic vector of all windows of single variable, The cluster obtaining will be clustered in magnitude order, maximum cluster is represented with character ' a ', secondary big cluster is represented with ' b ', the like, Then it is considered as noise less than the cluster of given threshold value 2, with '？' represent, by each window with its corresponding character representation of place cluster, so Raw value type data is converted to character string forms, i.e. the status word string of this variable；

4) multivariable state matrix generation module (1-4), by the status word string of all variables according to time unifying, forms state Matrix；

5) Apriori state relation excavate module (1-5) with Apriori algorithm, multivariable state matrix is carried out frequent item set and Association rule mining；

6) last, correlation rule yojan module (1-6) carries out yojan to the correlation rule detecting, and eliminates redundancy, obtains Final multivariable state relation rule.

2. according to claim 1 excavate time series data state relation method it is characterised in that：Described data is located in advance The step of reason module (1-1) elimination of burst noise includes：Calculate average and the standard deviation of each window, judge that each data point is located with it Whether the difference of watch window average is more than the standard deviation of 5 times of watch window, if being more than, this data point is unruly-value rejecting；Right Time series after elimination of burst noise carries out interpolation at equal intervals, if being spaced apart △ t, initial time is T, then time after interpolation at equal intervals Set should be { T+n* △ t n=0,1,2,3 ... }, and T+i* △ t is corresponding to be worth for nearest from this moment in original series Less than the value corresponding to T+i* △ t, that is, corresponding to first previous moment more than T+i* △ t in original series Observation；Linear normalization is carried out to the data after interpolation operation at equal intervals, scans a time series first, obtain observation The maximum (max) of value and minimum of a value (min), calculate the numerical value after the normalization of each observation station according to equation below, will be original Time series span is transformed on [0,1] interval；

$x_i=\frac{x_i-min}{\mathrm{\&Delta;}}$

Wherein, x_iRepresent i-th observation station numerical value；△=max-min.

3. the method excavating time series data state relation according to claim 1 is it is characterised in that characteristic extracting module (1-2)：First, with the window setting, univariate data is cut；Secondly, feature is carried out to the data in each window and carries Take, including statistical nature, Fourier's feature, wavelet character；V1=[average, variance], v1 represent statistical nature, and wherein average is anti- Reflect the average level of data in a window, variance then describes the degree of fluctuation of data in window；V2=[Fourier coefficient 1, Fourier frequency 1, Fourier coefficient 2, Fourier frequency 2], v2 represents frequency domain character, obtains one by Fourier transformation and is The Fourier coefficient of row, is ranked up according to absolute value from big to small to Fourier coefficient, chooses the maximum Fourier of the first two Coefficient and its corresponding frequency；V3=[wavelet coefficient detail coefficients 1 ... wavelet details coefficient n], v3 represent that time-frequency domain is special Levy, wavelet transform is carried out to each window, obtain n wavelet details coefficient, this tripartite's region feature is integrated, constitute Window feature vector v, v=[v1, v2, v3].

4. according to claim 1 excavate time series data state relation method it is characterised in that：Described dynamic division Cluster module (1-3) carries out clustering its process such as to the characteristic vector of all windows of single variable using dynamic partition clustering method Under：

1) the independent cluster of first window, the cluster heart is the multi-feature vector of this window；

2) the initial division process of cluster, calculates the similarity between the 2nd window and the 1st cluster cluster heart according to equation below：

$cos(v_{ik},v_{jk})=\frac{v_{ik}\&CenterDot;v_{jk}}{\left|v_{ik}\right|\&times;\left|v_{jk}\right|}$

In formula：cos(v_ik, v_jk) represent window i v_kThe v of (k=1,2,3) vector sum window j_kCosine similarity between vector； cos(v_ik, v_jk) ∈ [- 1 ,+1], using equation below, calculate the distance between two windows；

$d(v_{ik},v_{jk})=1-\frac{cos(v_{ik},v_{jk})+1}{2}$

$dist(v_i,v_j)=\frac{\underset{k=1}{\overset{3}{\&Sigma;}}dist(v_{ik},v_{jk})}{3}$

In formula：dist(v_i, v_j) represent the distance between window i and j, dist (v_i, v_j) ∈ [0,1]；

If dist<D, d=0.2), then No. 2 windows are incorporated to first cluster and the cluster heart changed immediately according to equation below：

${\mathrm{cv}}_k=\frac{\underset{i=1}{\overset{n}{\&Sigma;}}{v}_{ik}}{n},(k=1,2,3)$

In formula：cv_kRepresent cluster heart c k-th characteristic vector, it be equal to cluster in all k-th characteristic vectors of window average, one The cluster heart c of individual cluster is exactly cv₁, cv₂, cv₃Combination, that is, in cluster all window multi-feature vectors mean value；If dist >=d, The independent cluster of No. 2 windows, processes remaining window successively：Calculate between i-th window and all cluster cluster hearts having produced away from From, pick out the cluster c nearest with it apart from dist, if dist<D, i window is incorporated to cluster c, otherwise independent cluster；

3) the adjustment process of cluster：(i=1,2 ... m), and the cluster heart calculating it with all clusters, apart from dist, is chosen to take out No. i-th window Select the dist and its corresponding cluster c of minimum, if dist≤d, and window i not in cluster c, is then moved by window i from original cluster To cluster c；If dist≤d, and window i is in cluster c, then window i is not operated；If dist>D, then by window i from original cluster In remove, independent cluster；Repeat said process, until having processed all windows, calculate the cluster heart of all clusters.If there is a cluster The heart there occurs change, then repeat the adjustment process of cluster, i.e. step 3) until the cluster heart of all clusters no longer changes；If all cluster hearts are all Constant, then execution step 4)；

4) merging process of cluster：Calculate the cluster heart distance of any two cluster, select closest two cluster c_i, c_j, and its corresponding Apart from dist, if dist≤α, α=0.3), then merge cluster c_i, c_j, and calculate the cluster heart of new cluster after merging, repeat to merge Journey 4), if dist>α, represents that there are not two close enough clusters can merge, then exit merging process, clustering algorithm terminates, Window feature in same cluster in cluster result approximate it is considered to be a kind of state, the window of different clusters represents different shapes State；By all clusters in magnitude order, maximum cluster is represented with character ' a ', and secondary big cluster is represented with ' b ', the like, it is less than The cluster of given threshold value is then considered as noise, with '？' represent, by each window with its corresponding character representation of place cluster, so original Numeric type data is converted to character string forms, that is, obtain the status word string of each variable.

5. according to claim 1 excavate time series data state relation method it is characterised in that：Described multivariable shape State matrix generation module (1-4) is it is assumed that there being n variable by the status word string of all variables according to time unifying, each change The initial observation time of amount is identical with deadline, then their window number is necessarily identical, status word string length also phase With it is assumed that status word string length is m, then generate the multivariable state matrix of n*m.

6. according to claim 1 excavate time series data state relation method it is characterised in that：Described Apriori It is to carry out frequent item set and correlation rule using Apriori algorithm to multivariable state matrix that state relation excavates module (1-5) Excavate：Apriori algorithm Mining Frequent Itemsets Based flow process is as follows：First, by scanning transaction journal, find out all of frequent 1 Collection, this set is denoted as L₁, then utilize L₁Find frequent 2 item collections L₂, so on, until any frequent k item can not be found again Collection；Iteration is divided into two steps every time：First, produce candidate by connecting step and beta pruning step；Second, calculate each candidate The support of item, is considered as frequent episode more than the item of minimum support threshold value 0.001, excavates and close on the basis of frequent item set Connection rule, specific as follows：Firstly, for each frequent item set L, produce all nonvoid subsets of L；Secondly, each for L is non- Vacuous subset s, produces a candidate rule " s → (L-s) ", wherein (L-s) remaining content after representing removing s in L；Such as finally Really the confidence level of this candidate rule is more than given threshold value 0.5, then export this rule, otherwise abandons this rule, the confidence level of rule It is calculated as follows：

$Cf(L,s)=\frac{Sp\left(L\right)}{Sp\left(s\right)}$

Wherein, (L, s) represents the confidence level of regular " s → (L-s) " to Cf, and Sp (L) represents the support of L, and Sp (s) represents propping up of s Degree of holding.

7. according to claim 1 excavate time series data state relation method it is characterised in that：Described correlation rule Yojan module (1-6) merges to the redundancy rule producing or deletes, and reduction steps are：

1) for the correlation rule obtaining, it is ranked up from big to small according to confidence level；

2) for each K rank frequent episode (K>1), only retain the maximum correlation rule of confidence level；

3) if the former piece of two correlation rules is identical, consequent is compared, if consequent has inclusion relation, in confidence On the premise of degree difference very little, delete the rule belonging to consequent being comprised；

4) if the consequent of two correlation rules is identical, former piece is compared, if former piece has inclusion relation, in confidence On the premise of degree difference very little, delete the more rules of former piece, retain the fewer rule of former piece；

5) in order to ensure the uniformity of knowledge, it is to avoid circular reasoning occurs, need to detect with the presence or absence of ring in correlation rule, to ring Detection realized by directed acyclic graph, represent former piece with a node, a node represents consequent, the two is with directed edge even Connect, detect the correlation rule according to confidence level descending one by one.