CN105491614A - Wireless sensor network abnormal event detection method and system based on secondary mixed compression - Google Patents

Abstract

The invention discloses a wireless sensor network abnormal event detection method and system based on secondary mixed compression. The method comprises the steps that S1. primary compression is performed on original data sequences by a compression sensing method; and secondary compression is performed on the compressed sequences by a piecewise linear fitting method so that state edge operators are obtained; S2. edge amplitude and edge intensity of each data point in the compressed sequences are calculated, and new edge point sequences are formed by selecting edge points of low interpolation error from edge point sequences; S3. characteristic value sequences are obtained according to edge point sequences; S4. local accessible density and local abnormal factors are calculated; and S5. event detection is performed in corresponding time sequence intervals according to the size of the local abnormal factors after mixed compression. Search efficiency and abnormal data accurate positioning capacity of wireless sensor network nodes can be enhanced so that abnormal events can be more efficiently and rapidly found; and timeliness of abnormity detection can be enhanced so that energy and communication bandwidth can be greatly saved.

Description

Based on the wireless sense network accident detection method and system of secondary mixing compression

Technical field

The present invention relates to technology of Internet of things field, particularly relate to a kind of wireless sense network accident detection method and system based on secondary mixing compression.

Background technology

Wireless sense network (WirelessSensorNetworks is called for short WSN) is the dominant form of future network development, and has become new science research field in this century.Many urgent problems are proposed in basic theory and engineering two aspects.Wireless sense network is with low cost, low-power consumption, on a large scale MANET; Sensor node compact, powered battery, deployment are flexible; And can adapt to monitor manpower be difficult to arrive adverse circumstances; These features make wireless sense network greatly improve the monitoring capacity of disaster prevention.In order to monitor various contingent accident (as landslide, air pollution, forest fire etc.) in time, the abnormal measures that sensor node collects must be paid close attention to.Therefore, detect abnormal data real-time and accurately, and early warning particular event, tool is of great significance.

The accident detection technology of wireless sense network sums up and is mainly divided into two classes: 1) put method for detecting abnormality.If the abnormal i.e. sensing data of point exceedes certain threshold value of setting, then think that event occurs.This method is only suitable on a small scale, the single incident monitoring task of short-term.2) pattern method for detecting abnormality.In some long-term gradient environment monitorings, paroxysmal complicated event is often difficult to be reported to the police by transfiniting of specified attribute threshold value, can not describe by simple threshold value, but a kind of pattern (event schema) can be regarded as, carry out abnormality detection because mode identification technology can be adopted.At present, most of pattern method for detecting abnormality is all carry out on acquired original data space, does not namely carry out any conversion to the data of sensor node collection, although this method has certain accuracy of detection.But algorithm amount of calculation is large, poor fault tolerance, and energy-saving effect is limited.Abnormality detection can be carried out on the data space after overcompression process.Further, in wireless sense network, accident detection technology also will face two significant challenge: 1) accuracy of detection.Due to the impact by various fault in ambient noise and network, sensor node is often to the monitor value made mistake, and this will certainly have influence on the reliability of accident detection.Therefore, detection method must have fault-tolerance.2) energy efficiency.Sensor node has very limited energy reserve, and the network lifetime of wireless sense network event monitoring depends on node energy consumption, and thus detection method must have energy saving.

Extensive long-term Dimension Time Series of disposing thousands of sensor node generation magnanimity in wireless sense network, contain a large amount of redundancies in these data and conceal the correlation of important relationship, if directly carry out abnormality detection in these original data space, the great expense incurred of its energy and communication bandwidth will shorten network lifecycle, even makes wireless sense network can not complete monitoring task.Therefore, before data are sent to gateway, carry out compressing (or dimensionality reduction) to be very important.In event monitoring type WSN application system, from the Monitoring Data of network, identify that anomalous event is its primary goal fast, its importance even exceedes Monitoring Data itself.The temporal correlation between node is excavated by data compression method, eliminate the redundancy between data to greatest extent, while guarantee significantly reduces volume of transmitted data, still can keep high-precision event monitoring performance, and extract potential useful information, pattern and trend from magnanimity flow data.

Succinctly, flexibly, represent that signal is that field of information processing is the most basic always adaptively, the studying a question of forefront.Along with people constantly increase the signal bandwidth demand of carry information, based on classical signal sampling thheorem, hardware A/D is sampled and processing speed also more and more faster, thus hard to carry on to the Wideband Signal Processing.Such as, the multispectral environmental resource exploration of high accuracy, its mass data transfers and process are exactly a difficult problem.But " Nyquist " sampling rate must reach more than the twice of signal bandwidth and just accurately reconstruct the abundant of primary signal but not necessary condition.In recent years, the compressive sensing theory that the people such as Donoho, Candes and Tao propose has broken the constraint of classical " Shannon/Nyquist " sampling thheorem, as long as the information processing framework of compressive sensing theory shows that signal has openness (or compressibility), just can go sampled signal with far below Nyquist rate, and from a small amount of sampled value, ideally can recover primary signal with high probability, without the need to considering that the concrete physics of signal estimates the restriction of (as frequency, bandwidth).Therefore, this theory is that the design of wireless sense network data compression method provides new way.Combining wireless Sensor Network relevant feature, the data dependence between node is excavated by compression sensing method, eliminate the redundancy between data to greatest extent, while guarantee significantly reduces volume of transmitted data, still can keep high-precision event monitoring performance, and extract potential useful information, pattern and trend from magnanimity flow data.

Summary of the invention

The technical problem to be solved in the present invention be in prior art to the magnanimity height latitude time series data that a large amount of sensor node produces, be difficult to the defect of directly carrying out abnormality detection, there is provided a kind of to compress and dimensionality reduction initial data, reduce the wireless sense network accident detection method and system based on secondary mixing compression of communication bandwidth.

The technical solution adopted for the present invention to solve the technical problems is:

The invention provides a kind of wireless sense network accident detection method based on secondary mixing compression, comprise the following steps:

S1, slip detection window is set, by compression sensing method, first time compression is carried out to the original data sequence X gathered in sliding window, obtain compressed sequence Y; And by sectional linear fitting method, second time compression is carried out to compressed sequence Y, obtain the expansion tense boundary operator of each data point in compressed sequence;

S2, calculate each data point edge amplitude e in compressed sequence Y according to expansion tense boundary operator _iwith edge strength R _i, select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y, and in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

S3, according to marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection; And computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Any two pattern c in S4, calculating C (f) _iand c _jpattern distance dist (c _i, c _j), computation schema f _ithis locality can arrive density lrd _k(f _i) and pattern f _ilocal Outlier factor LOF _k(f _i);

S5, according to mixing compression after local Outlier factor size carry out event detection in the time series interval of correspondence.

Further, carrying out to original data sequence X the formula that first time compression obtains compressed sequence Y in step S1 of the present invention is:

Wherein, original series X ∈ R ⁿ, compressed sequence Y ∈ R ^m, the dimension m<<n of Y, φ ^{m × n}for the sparse matrix of signal, ψ ^{n × n}for perception matrix, and sparse matrix and perception matrix meet RIP character, according to the time fluctuation feature of original series X, select corresponding sparse matrix ψ ^{n × n}.

Further, expansion tense boundary operator ETEO (t, w) of compressed sequence Y is calculated in step S1 of the present invention _iformula be:

ETEO(t,w) _i＝(y _i+t-y _i)

Wherein, 1≤i≤m ,-w≤t≤w, i is expansion tense boundary operator detection window length is the central point of w.

Further, each data point edge amplitude e in compressed sequence Y is calculated in step S2 of the present invention _iwith edge strength R _iformula be respectively:

$e_i=\mathrm{\&Sigma;}(y_{i+t}*ETEO(t,w\left)\right)$

$R_i=\underset{k=i-w}{\overset{k=i+w}{\&Sigma;}}p\left(k\right),k\&NotEqual;i$

Wherein ， ﹡ represents discrete convolution, and in formula, p (k) is defined as follows:

$p\left(k\right)=\left\{\begin{array}{c}1,e_i-e_k>0\\ -1,e_i-e_k<0\end{array}\right.$

Node adopts expansion tense boundary operator and compressed sequence Y to carry out convolution algorithm, obtains each data point y _iedge amplitude edge amplitude e _iwith edge strength R _i.

Further, set of patterns Y is calculated in step S3 of the present invention _lmiddle f _ithe characteristic length L of pattern _i, slope K _iwith average M _iformula be respectively: L _i=t _i+1-t _i+ 1,

Further, in step S3 of the present invention by f _ithe formula that the length of pattern feature, slope, average are normalized is respectively:

$l_i=\frac{L_i-min\left(L_i\right)}{\max \left(L_i\right)-\min \left(L_i\right)}$

$k_i=\frac{K_i-min\left(K_i\right)}{\max \left(K_i\right)-\min \left(K_i\right)}$

$m_i=\frac{M_i-min\left(M_i\right)}{max\left(M_i\right)-\min \left(M_i\right)}$

Wherein, l _i, k _iand m _irepresent length, slope and the average after normalized respectively.

Further, any two pattern c in C (f) are calculated in step S4 of the present invention _iand c _jpattern distance dist (c _i, c _j) formula is: $dist(c_i,c_j)=\sqrt{(l_i-l_j)+(k_i-k_j)+(m_i-m_j)}.$

Further, computation schema f in step S4 of the present invention _ithis locality can arrive density lrd _k(f _i) formula is:

${\mathrm{lrd}}_k\left(f_i\right)=\frac{\underset{f_i\&Element;N_k\left(f_i\right)}{\&Sigma;}rd(c_i,c_j)}{\left|N_k\left(f_i\right)\right|}$

Wherein, given k ∈ N ⁺, N _k(f _i) represent and pattern f _idistance be not more than k_dist (f _i) all set of modes, be called pattern f _ithe nearest neighbour of k, | N _k(f _i) | represent the number of modes in set;

Rd (c _i, c _j) intermediate scheme c _irelative to c _jarrived in distance, rd (c _i, c _j) computing formula be: rd (c _i, c _j)=max (k_dist (c _j), dist (c _i, c _j)).

Further, computation schema f in step S4 of the present invention _ilocal Outlier factor LOF _k(f _i) formula is:

${\mathrm{LOF}}_k\left(f_i\right)=\frac{\underset{f_j\&Element;N_k\left(f_i\right)}{\&Sigma;}\frac{{\mathrm{lrd}}_k\left(f_i\right)}{{\mathrm{lrd}}_k\left(f_j\right)}}{\left|N_k\left(f_i\right)\right|}$

Wherein, lrd _k(f _i) be pattern f _ithis locality can arrive density, N _k(f _i) intermediate scheme f _ithe nearest neighbour of k.

The invention provides a kind of wireless sense network accident detection system based on secondary mixing compression, comprising:

Second-compressed unit, for arranging slip detection window, carrying out first time compression by compression sensing method to the original data sequence X gathered in sliding window, obtaining compressed sequence Y; And by sectional linear fitting method, second time compression is carried out to compressed sequence Y, obtain the expansion tense boundary operator of each data point in compressed sequence;

Border sequences computing unit, for calculating each data point edge amplitude e in compressed sequence Y according to expansion tense boundary operator _iwith edge strength R _i, select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y, and in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

Characteristic value sequence computing unit, for according to marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection; And computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Local Outlier factor computing unit, for calculating any two pattern c in C (f) _iand c _jpattern distance dist (c _i, c _j), computation schema f _ithis locality can arrive density lrd _k(f _i) and pattern f _ilocal Outlier factor LOF _k(f _i);

Event detection unit, for carrying out event detection according to the local Outlier factor size after mixing compression in the time series interval of correspondence.

The beneficial effect that the present invention produces is: the wireless sense network accident detection method based on secondary mixing compression of the present invention, by greatly reducing the amount of calculation of wireless sense network abnormality detection in conjunction with compressed sensing technology, improve the ability of wireless sensing net node search efficiency and accurate location abnormal data, can more efficiently note abnormalities event rapidly;

By the data space of each for wireless sense network Node distribution after mixing compression is carried out abnormality detection, only abnormal information detected, and by after the detection misarrangement of spatial correlation, node could send to gateway the compressed sequence comprising abnormal information, improve the promptness of abnormality detection, significantly save energy and communication bandwidth.

Accompanying drawing explanation

Below in conjunction with drawings and Examples, the invention will be further described, in accompanying drawing:

Fig. 1 is the flow chart of the wireless sense network accident detection method based on secondary mixing compression of the embodiment of the present invention;

Fig. 2 is that the sectional linear fitting compression method of the wireless sense network accident detection method based on secondary mixing compression of the embodiment of the present invention detects X ₃the Experimental comparison figure of the local Outlier factor of (t);

Fig. 3 be the compressed sensing of wireless sense network accident detection method based on secondary mixing compression of the embodiment of the present invention mix with sectional linear fitting secondary compress after detect X ₃the local Outlier factor Experimental comparison figure of (t);

Fig. 4 is the structured flowchart of the wireless sense network accident detection system based on secondary mixing compression of the embodiment of the present invention;

In figure, 401-second-compressed unit, 402-border sequences computing unit, 403-characteristic value sequence computing unit, the local Outlier factor computing unit of 404-, 405-event detection unit.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

As shown in Figure 1, the wireless sense network accident detection method based on secondary mixing compression of the embodiment of the present invention, comprises the following steps:

S1, slip detection window is set, by compression sensing method, first time compression is carried out to the original data sequence X gathered in sliding window, obtain compressed sequence Y; And by sectional linear fitting method, second time compression is carried out to compressed sequence Y, obtain the expansion tense boundary operator of each data point in compressed sequence;

S2, calculate each data point edge amplitude e in compressed sequence Y according to expansion tense boundary operator _iwith edge strength R _i, select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y, and in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

S3, according to marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection; And computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Any two pattern c in S4, calculating C (f) _iand c _jpattern distance dist (c _i, c _j), computation schema f _ithis locality can arrive density lrd _k(f _i) and pattern f _ilocal Outlier factor LOF _k(f _i);

S5, according to mixing compression after local Outlier factor size carry out event detection in the time series interval of correspondence.

First sliding window is set at wireless sensing net node, utilizes compressed sensing technology to the original data sequence X ∈ R gathered in sliding window ⁿcarry out first time compression, obtain compressed sequence Y:

Y＝φψ ^TX(1)

Wherein, Y ∈ R ^m, the dimension m<<n of Y, as long as according to the sparse matrix φ of compressive sensing theory signal ^{m × n}with perception matrix ψ ^{n × n}meet RIP character, just can by sequence Y Accurate Reconstruction original series X, namely can not abnormal information in undetected original data space X, according to the time fluctuation feature of sequence X, corresponding sparse matrix ψ can be selected ^{n × n}, the amount of calculation of compression process is approximately o (mn) for the first time.

In order to reduce the search volume of abnormal data further, recycling sectional linear fitting method carries out second time compression to compressed sequence Y, obtains the waypoint that sectional linear fitting error is minimum, carrys out linear fit compressed sequence Y with the straightway of these waypoints; Calculate expansion tense boundary operator ETEO (t, w) of compressed sequence Y _iformula be:

ETEO(t,w) _i＝(y _i+t-y _i)(2)

Wherein, 1≤i≤m ,-w≤t≤w, i is expansion tense boundary operator detection window length is the central point of w, calculates the boundary operator of each data point in compressed sequence Y successively, obtains the expansion tense boundary operator sequence corresponding to data point.

According to expansion tense boundary operator, each data point edge amplitude e in definition compressed sequence Y _iformula is:

e _i＝Σ(y _i+t*ETEO(t，w))(3)

Wherein ， ﹡ represents discrete convolution, and node adopts expansion tense boundary operator and compressed sequence Y to carry out convolution algorithm, obtains each data point y _iedge amplitude edge amplitude e _i.Edge amplitude e _irepresent the variation tendency of sequence in this data point neighborhood, near larger then this point of edge amplitude, data fluctuations Shaoxing opera is strong; Less edge amplitude represents that the data point in its neighborhood is in same variation tendency substantially.

Sequence due to node collection is the random process becoming non-stationary for the moment.As time goes on, data characteristics may change, thus impact is to the judgement of sequence data marginal point with choose.For this reason, by edge amplitude e _icarry out digital quantization and obtain edge strength R _i, choose the large data point of edge strength as marginal point, the point that edge strength is larger, the possibility becoming marginal point is larger; When edge strength is identical, choose the less point of interpolation error as marginal point.Edge strength R _ibe defined as:

$R_i=\underset{k=i-w}{\overset{k=i+w}{\&Sigma;}}p\left(k\right),k\&NotEqual;i---\left(4\right)$

Wherein, p (k) is defined as follows:

$p\left(k\right)=\left\{\begin{array}{c}1,e_i-e_k>0\\ -1,e_i-e_k<0\end{array}\right.---\left(5\right)$

Select the edge strength R in detection window w _iextreme point as marginal point, choose the less N number of marginal point of interpolation error and form marginal point sequence utilize marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection, then the Piecewise Linear Representation of compressed sequence Y is claim Y _lin any one f _ifor an ETEO pattern of compressed sequence Y.Piecewise linearity compression process amount of calculation is o (γ logm), and wherein γ is the straight segments number of sequence Y, generally speaking γ < < m.

Whether the data sequence of node collection has exception can be reflected in corresponding pattern feature has without exception, and when pattern feature occurs abnormal, this pattern is likely abnormal patterns.Abnormal patterns is usually expressed as a kind of exception of pattern feature or the comprehensive of various modes feature abnormalities; therefore; for comprehensively representing data sequence abnormal patterns feature; choose the length of ETEO pattern, slope and average as pattern feature; data sequence is mapped to three-dimensional feature space, Three models feature is defined as follows:

If f _icompressed sequence Y ETEO pattern, definition f _ithe characteristic length L of pattern _i, slope K _iwith average M _ifor:

$L_i=t_{i+1}-t_i+1,K_i=\frac{y_{t_{i+1}}-y_{t_i}}{L_i-1},M_i=\frac{y_{t_{i+1}}+y_{t_i}}{2}---\left(6\right)$

Further, in order to obtain f _icharacteristic value sequence C (f)={ c of pattern ₁, c ₂..., c _γ, wherein c _i=(l _i, k _i, m _i) intermediate scheme f _ifeature.Due to f _ithe length of pattern feature, slope, average codomain are different, three kinds of characteristic values are adopted extreme difference method for normalizing Linear Mapping to interval [0,1], carry out standardization processing formula and be respectively:

$l_i=\frac{L_i-\min \left(L_i\right)}{\max \left(L_i\right)-\min \left(L_i\right)},k_i=\frac{K_i-\min \left(K_i\right)}{\max \left(K_i\right)-\min \left(K_i\right)},m_i=\frac{M_i-\min \left(M_i\right)}{\max \left(M_i\right)-\min \left(M_i\right)}---\left(7\right)$

Land use models feature c _iwith Euclidean distance measures the distance between two different modes, any two pattern c in definition C (f) _iand c _jpattern distance dist (c _i, c _j) be:

$dist(c_i,c_j)=\sqrt{(l_i-l_j)+(k_i-k_j)+(m_i-m_j)}---\left(8\right)$

Given k ∈ N ⁺(positive integer collection), defining mode f _ithis locality can arrive density lrd _k(f _i) be:

${\mathrm{lrd}}_k\left(f_i\right)=\frac{\underset{f_i\&Element;N_k\left(f_i\right)}{\&Sigma;}rd(c_i,c_j)}{\left|N_k\left(f_i\right)\right|}---\left(9\right)$

Wherein, N _k(f _i) represent and pattern f _idistance be not more than k_dist (f _i) all set of modes, be called pattern f _ithe nearest neighbour of k, | N _k(f _i) | represent the number of modes in set, to appointing pattern f _ik distance definition be from f _ithe distance of a kth pattern of arest neighbors, its computing formula is:

k_dist(f _i)＝dist(c _i,c _j)(10)

Due to multiple pattern and pattern f may be had _idistance equal k_dist (f _i), therefore k_dist (f _i) should meet the following conditions simultaneously:

1) k pattern f is had at least _j, and j ≠ i, meet dist (c _i, c _j)≤k_dist (f _i);

2) k-1 pattern f is had at the most _j, and j ≠ i, meet dist (c _i, c _j) < k_dist (f _i);

Rd (c _i, c _j) be called pattern c _irelative to c _jarrived in distance, its computing formula is:

rd(c _i,c _j)＝max(k_dist(c _j),dist(c _i,c _j))(11)

This locality can arrive density lrd _k(f _i) have expressed pattern f _ineighborhood density distribution, given k ∈ N ⁺, defining mode f _ilocal Outlier factor LOF _k(f _i) be:

${\mathrm{LOF}}_k\left(f_i\right)=\frac{\underset{f_j\&Element;N_k\left(f_i\right)}{\&Sigma;}\frac{{\mathrm{lrd}}_k\left(f_i\right)}{{\mathrm{lrd}}_k\left(f_j\right)}}{\left|N_k\left(f_i\right)\right|}---\left(12\right)$

F _ithe local Outlier factor LOF of pattern _k(f _i) illustrating pattern intensity of anomaly, Outlier factor is larger, and this pattern is got over modes of departure and to be clustered center, larger with other pattern differentials, when Outlier factor exceedes a certain set-point, think that this pattern is abnormal, wireless sense network judges that sequence of interval corresponding to this pattern exists anomalous event.

Adopt local outlier factor LOF _k(f _i) amount of calculation of carrying out abnormality detection is made up of three parts: (1) computation schema distance is o (γ ²), (2) computation schema local reachability density lrd _k(f _i) be o (γ k), wherein k is number of modes in neighborhood and k≤γ, and (3) calculate local outlier factor LOF _k(f _i) be also o (γ k), then the amount of calculation of abnormality detection is approximately o (γ ²).Due to γ < < m < < n, therefore, the method greatly reduces the computation complexity of wireless sense network abnormality detection.

According to above related definition and theory analysis, the present invention proposes a kind of method for quick of the wireless sense network monitoring anomalous event based on secondary mixing compression.First the method utilizes compressed sensing technology that the initial data of wireless sense network collection is carried out dimensionality reduction, recycling sectional linear fitting method carries out second time compression to the sequence after dimensionality reduction, obtain the waypoint that sectional linear fitting error is minimum, linear fit sequence is carried out with the straightway of these waypoints, local outlier factor is adopted to carry out abnormality detection again, local Outlier factor in mode characterizes the mixing compressed intensity of anomaly according to sequence of secondary, instead of directly on original data sequence, carries out abnormality detection to individual data point.

In another embodiment of the present invention, be input as and arrange wireless sensing net node detection sliding window length, export as whether wireless sense network report monitored area has anomalous event, the concrete steps of the method are:

Step one: utilize compressed sensing technology to the original data sequence X ∈ R gathered in sliding window ⁿcarry out first time compression, obtain compressed sequence Y=φ ψ ^tx, (Y ∈ R ^m);

Step 2: utilize sectional linear fitting method to carry out second time compression to compressed sequence Y, calculates each data point expansion tense boundary operator ETEO (t, w) in Y _i;

Step 3: calculate each data point edge amplitude e in compressed sequence Y _iwith edge strength R _i;

Step 4: select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y;

Step 5: in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

Step 6: utilize marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection;

Step 7: computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Step 8: calculate any two pattern c in C (f) _iand c _jpattern distance dist (c _i, c _j);

Step 9: computation schema f _ithis locality can arrive density lrd _k(f _i);

Step 10: computation schema f _ilocal Outlier factor LOF _k(f _i);

Step 11: wireless sensing net node carries out event detection according to the local Outlier factor size after secondary mixing compression in the time series interval of correspondence.

Verify by experiment, in experiment, adopt disclosed generated data collection Ma_Data to be produced by following process:

$X_1\left(t\right)=sin\left(\frac{40\mathrm{\&pi;}}{n}t\right)+n\left(t\right)$

$X_2\left(t\right)=sin\left(\frac{40\mathrm{\&pi;}}{n}t\right)+n\left(t\right)+e_1\left(t\right)$

$X_3\left(t\right)=\sin \left(\frac{40\mathrm{\&pi;}}{n}t\right)+n\left(t\right)+e_1\left(t\right)+e_2\left(t\right)$

Wherein, t=1,2 ..., n, n=1200, n (t) are averages is 0, and standard deviation is 0.1 additive Gaussian noise, e ₁(t), e ₂t () is two anomalous events, e ₁t () is defined as follows:

N ₁t () meets normal distribution N (0,0.5), e ₂t () is defined as follows:

From defining above, data sequence X ₁(t) to be the cycle be 60 normal sequence, X ₂t () is at X ₁abnormal e is with the addition of in (t) ₁(t), X ₃t () is at X ₂abnormal e is with the addition of in (t) ₂t (), changing the character of sequence in interval [600,620], [820,870], there is anomalous event in these two intervals.

As shown in Figures 2 and 3, carry out contrast experiment to evaluate performance of the present invention, Fig. 2 only adopts sectional linear fitting compression method to detect X ₃t the local Outlier factor of (), Fig. 3 adopts compressed sensing to detect X with sectional linear fitting secondary mixing compression method ₃the local Outlier factor of (t).

In figs. 2 and 3, the first half curve represents the sequence X on Ma_Data data set ₃t (), the latter half curve represents the Outlier factor of corresponding data sequence.Outlier factor in Fig. 2 between interval [600,620] is interval apparently higher than other, accurately detects anomalous event e ₁(t); The Outlier factor of interval [800,870] is also comparatively large, and also accurately note abnormalities event e simultaneously ₂(t).

X in Fig. 3 ₃t () first carries out first time compression through compressed sensing technology, sliding window length is set to 30, obtaining compressed sequence Y length is 10, after recycling sectional linear fitting carries out second time compression, Fig. 3 sequence of interval length reduces 3 times compared with Fig. 2, local outlier factor interval near data point 200 and 275 is all interval than other large, and not only accurately note abnormalities event e compared with Fig. 2 ₁t (), also accurately note abnormalities event e simultaneously ₂(t), and detect anomalous event e ₂t the local outlier factor of () is compared Fig. 2 and is contrasted more obvious.The results show, after secondary mixing compression, greatly reduces the spatial dimension of search abnormal data, focuses on and highlights the position of abnormal data.

In sum, the present invention proposes a kind of method for quick of the wireless sense network monitoring anomalous event based on secondary mixing compression.First the method utilizes compressed sensing technology that the initial data of wireless sense network collection is carried out dimensionality reduction, recycling sectional linear fitting method carries out second time compression to the sequence after dimensionality reduction, obtain the waypoint that sectional linear fitting error is minimum, linear fit sequence is carried out with the straightway of these waypoints, local outlier factor is adopted to carry out abnormality detection again, by the amount of calculation of abnormality detection from o (n ²)) be reduced to o (γ ²), improve the ability of node searching efficiency and accurate location abnormal data.Experimental result shows, and directly carries out compared with abnormality detection in original data space, and more can efficiently note abnormalities event fast.

As shown in Figure 4, the wireless sense network accident detection system based on secondary mixing compression of the embodiment of the present invention, for realizing the wireless sense network accident detection method based on secondary mixing compression of the embodiment of the present invention, comprising:

Second-compressed unit 401, for arranging slip detection window, carrying out first time compression by compression sensing method to the original data sequence X gathered in sliding window, obtaining compressed sequence Y; And by sectional linear fitting method, second time compression is carried out to compressed sequence Y, obtain the expansion tense boundary operator of each data point in compressed sequence;

Border sequences computing unit 402, for calculating each data point edge amplitude e in compressed sequence Y according to expansion tense boundary operator _iwith edge strength R _i, select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y, and in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

Characteristic value sequence computing unit 403, for according to marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection; And computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Local Outlier factor computing unit 404, for calculating any two pattern c in C (f) _iand c _jpattern distance dist (c _i, c _j), computation schema f _ithis locality can arrive density lrd _k(f _i) and pattern f _ilocal Outlier factor LOF _k(f _i);

Event detection unit 405, for carrying out event detection according to the local Outlier factor size after mixing compression in the time series interval of correspondence.

Should be understood that, for those of ordinary skills, can be improved according to the above description or convert, and all these improve and convert the protection range that all should belong to claims of the present invention.

Claims (10)

1., based on a wireless sense network accident detection method for secondary mixing compression, it is characterized in that, comprise the following steps:

S1, slip detection window is set, by compression sensing method, first time compression is carried out to the original data sequence X gathered in sliding window, obtain compressed sequence Y; And by sectional linear fitting method, second time compression is carried out to compressed sequence Y, obtain the expansion tense boundary operator of each data point in compressed sequence;

S2, calculate each data point edge amplitude e in compressed sequence Y according to expansion tense boundary operator _iwith edge strength R _i, select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y, and in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

S3, according to marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection; And computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Any two pattern c in S4, calculating C (f) _iand c _jpattern distance dist (c _i, c _j), computation schema f _ithis locality can arrive density lrd _k(f _i) and pattern f _ilocal Outlier factor LOF _k(f _i);

S5, according to mixing compression after local Outlier factor size carry out event detection in the time series interval of correspondence.

2. the wireless sense network accident detection method based on secondary mixing compression according to claim 1, is characterized in that, carries out compressing the formula obtaining compressed sequence Y being for the first time in step S1 to original data sequence X:

Wherein, original series X ∈ R ⁿ, compressed sequence Y ∈ R ^m, the dimension m<<n of Y, φ ^{m × n}for the sparse matrix of signal, ψ ^{n × n}for perception matrix, and sparse matrix and perception matrix meet RIP character, according to the time fluctuation feature of original series X, select corresponding sparse matrix ψ ^{n × n}.

3. the wireless sense network accident detection method based on secondary mixing compression according to claim 1, is characterized in that, calculate expansion tense boundary operator ETEO (t, w) of compressed sequence Y in step S1 _iformula be:

ETEO(t,w) _i＝(y _i+t-y _i)

Wherein, 1≤i≤m ,-w≤t≤w, i is expansion tense boundary operator detection window length is the central point of w.

4. the wireless sense network accident detection method based on secondary mixing compression according to claim 3, is characterized in that, calculate each data point edge amplitude e in compressed sequence Y in step S2 _iwith edge strength R _iformula be respectively:

e _i＝∑(y _i+t*ETEO(t,w))

$R_i=\underset{k=i-w}{\overset{k=i+w}{\&Sigma;}}p\left(k\right),k\&NotEqual;i$

Wherein ， ﹡ represents discrete convolution, and in formula, p (k) is defined as follows:

$p\left(k\right)=\left\{\begin{array}{c}1,e_i-e_k>0\\ -1,e_i-e_k<0\end{array}\right.$

Node adopts expansion tense boundary operator and compressed sequence Y to carry out convolution algorithm, obtains each data point y _iedge amplitude edge amplitude e _iwith edge strength R _i.

5. the wireless sense network accident detection method based on secondary mixing compression according to claim 1, is characterized in that, calculate set of patterns Y in step S3 _lmiddle f _ithe characteristic length L of pattern _i, slope K _iwith average M _iformula be respectively: L _i=t _i+1-t _i+ 1,

6. the wireless sense network accident detection method based on secondary mixing compression according to claim 5, is characterized in that, by f in step S3 _ithe formula that the length of pattern feature, slope, average are normalized is respectively:

$l_i=\frac{L_i-min\left(L_i\right)}{\max \left(L_i\right)-\min \left(L_i\right)}$

$k_i=\frac{K_i-min\left(K_i\right)}{\max \left(K_i\right)-\min \left(K_i\right)}$

$m_i=\frac{M_i-min\left(M_i\right)}{max\left(M_i\right)-\min \left(M_i\right)}$

Wherein, l _i, k _iand m _irepresent length, slope and the average after normalized respectively.

7. the wireless sense network accident detection method based on secondary mixing compression according to claim 6, is characterized in that, calculate any two pattern c in C (f) in step S4 _iand c _jpattern distance dist (c _i, c _j) formula is: $dist(c_i,c_j)=\sqrt{(l_i-l_j)+(k_i-k_j)+(m_i-m_j)}.$

8. the wireless sense network accident detection method based on secondary mixing compression according to claim 7, is characterized in that, computation schema f in step S4 _ithis locality can arrive density lrd _k(f _i) formula is:

${\mathrm{lrd}}_k\left(f_i\right)=\frac{\underset{f_i\&Element;N_k\left(f_i\right)}{\&Sigma;}rd(c_i,c_j)}{\left|N_k\left(f_i\right)\right|}$

Wherein, given k ∈ N ⁺, N _k(f _i) represent and pattern f _idistance be not more than k_dist (f _i) all set of modes, be called pattern f _ithe nearest neighbour of k, | N _k(f _i) | represent the number of modes in set;

Rd (c _i, c _j) intermediate scheme c _irelative to c _jarrived in distance, rd (c _i, c _j) computing formula be: rd (c _i, c _j)=max (k_dist (c _j), dist (c _i, c _j)).

9. the wireless sense network accident detection method based on secondary mixing compression according to claim 8, is characterized in that, computation schema f in step S4 _ilocal Outlier factor LOF _k(f _i) formula is:

${\mathrm{LOF}}_k\left(f_i\right)=\frac{\underset{f_j\&Element;N_k\left(f_i\right)}{\&Sigma;}\frac{{\mathrm{lrd}}_k\left(f_i\right)}{{\mathrm{lrd}}_k\left(f_j\right)}}{\left|N_k\left(f_i\right)\right|}$

Wherein, lrd _k(f _i) be pattern f _ithis locality can arrive density, N _k(f _i) intermediate scheme f _ithe nearest neighbour of k.

10., based on a wireless sense network accident detection system for secondary mixing compression, it is characterized in that, comprising:

Second-compressed unit, for arranging slip detection window, carrying out first time compression by compression sensing method to the original data sequence X gathered in sliding window, obtaining compressed sequence Y; And by sectional linear fitting method, second time compression is carried out to compressed sequence Y, obtain the expansion tense boundary operator of each data point in compressed sequence;

Border sequences computing unit, for calculating each data point edge amplitude e in compressed sequence Y according to expansion tense boundary operator _iwith edge strength R _i, select the edge strength R in detection window w _iextreme point as marginal point, join marginal point sequence C _y, and in marginal point sequence C _yin choose the less N number of marginal point of interpolation error, form new marginal point sequence

Characteristic value sequence computing unit, for according to marginal point sequence be Y by compressed sequence Y Piecewise Linear Representation _l, by Y _lin all f _icompositional model collection; And computation schema concentrates f _ithe characteristic length of pattern, slope and average, and carry out standardization processing, obtain characteristic value sequence: C (f)={ c ₁, c ₂..., c _γ;

Local Outlier factor computing unit, for calculating any two pattern c in C (f) _iand c _jpattern distance dist (c _i, c _j), computation schema f _ithis locality can arrive density lrd _k(f _i) and pattern f _ilocal Outlier factor LOF _k(f _i);

Event detection unit, for carrying out event detection according to the local Outlier factor size after mixing compression in the time series interval of correspondence.