CN110569890A - A method for detecting abnormal patterns in hydrological data based on similarity measure - Google Patents

Abstract

The invention discloses a method for detecting abnormal patterns of hydrological data based on similarity measurement. The method is based on the linear segment representation of key points and represents KPRA-PLR algorithm, cuts hydrological data according to the definition of key points, and performs PLR algorithm for each subsequence. Line fitting is performed, and the subsequence is represented by the slope a _i of the straight line and the time interval _Δt ; each subsequence after segmentation is called a meta-pattern, and the sequence pattern is obtained by combining adjacent meta-patterns. Using the weighted distance and SDTW algorithm respectively, the similarity measure of each sequence pattern is calculated by using the weighted distance and the SDTW algorithm, and then calculating the anomaly score of each sequence pattern, which is the inverse of the average distance between the pattern and other patterns; the anomaly score is the k-nearest neighbor distance of the sequence pattern S _x , according to k- The nearest neighbor local detection principle calculates the local outlier factor LOF. The abnormal pattern detected by the method of similarity measurement in the present invention is more accurate, and a new technology is provided for the abnormal pattern detection of hydrological data from the perspective of data analysis.

Description

A method for detecting abnormal patterns in hydrological data based on similarity measure

technical field

The invention belongs to the hydrological detection technology, in particular to a method for detecting abnormal patterns of hydrological data based on similarity measurement.

Background technique

Hydrological time series is a series of observed values of physical quantities (water level, flow, rainfall, etc.) obtained by the observation system in time sequence. Hydrological time series is a common complex data type, which objectively records the observation information obtained by the observation system in chronological order. In my country, with the development of water conservancy informatization, the hydrological time series will be transmitted to the water conservancy information sharing platform, processed by the staff, and stored in the national hydrological library. Due to the measurement error of the acquisition equipment, the error of manual operation, and the change of the evolution law of the hydrology itself, there will be some abnormal data in the hydrology time series.

The main goal of Anomaly Detection is to mine unusual data hidden in large data sets, that is, potentially meaningful information that is significantly different from other data. However, the methods and technologies for detecting abnormal data in hydrological time series are still in the exploratory stage, and there is no perfect anomaly detection solution in the industry. At present, the methods for detecting abnormal points in hydrological time series include Box-Plot method and Benford's rule.

SUMMARY OF THE INVENTION

Purpose of the invention: Aiming at the problems of low accuracy and slow detection of abnormal conditions in the prior art, the present invention provides a method for detecting abnormal patterns in hydrological data based on similarity measurement, which provides a theoretical basis for detecting abnormal patterns in hydrological data. .

Technical solution: A method for detecting anomalous patterns in hydrological data based on similarity measurement, the method is based on the piecewise linear representation of key points KPRA-PLR algorithm, a method for similarity measurement of time series patterns, and local anomaly detection based on k-nearest neighbors The algorithm includes the following steps:

(1) KPRA-PLR algorithm based on piecewise linear representation of key points: First, the original hydrological time series is divided according to the definition of key points to obtain each subsequence, and then each subsequence is fitted with a straight line according to the PLR algorithm, Obtain the slope of the straight line, solve the intersection between two adjacent straight lines, and obtain the start time t _il and end time t _ir of each subsequence, and use ( _ai , t _il , t _ir ) to represent each subsequence, also known as the element model;

(2) Method for measuring the similarity of time series patterns: combine adjacent meta-patterns into sequential patterns; when measuring the similarity of two patterns, establish a similarity distance measure function between the two patterns, which is expressed by the distance The similarity between the two, the average similarity of each pattern represents the abnormal score of the pattern;

(3) k-nearest neighbor local anomaly detection algorithm: the k-nearest neighbor distance of the model is represented by the anomaly score value, and the k-nearest neighbor local reachable distance, local reachability density, and local anomaly factor are obtained. The largest anomaly factor is the anomaly. model.

Further, step (1) includes the following process:

(11) Define the key point as S={s ₁ , s ₂ , ..., s _m } as all local extreme points of the time series X, when the extreme point _sk satisfies |s _k -s _k-1 When |<ε and |t _k -t _k-1 |<δ, where ε and δ are given constants, and t _k represents the observation time of each hydrological physical quantity, then s _k is called an unimportant local extreme point. Other extreme points in S are key points;

(12) Then, according to the linear segment representation algorithm, use the least squares method to fit a straight line through two adjacent key points to obtain the slope corresponding to the straight line, solve the intersection between the two adjacent straight lines, and obtain the starting time of each subsequence and the termination moments t _il and t _ir ; the pattern of each subsequence is represented as ( _ai , t _il , t _ir ) is called a meta-pattern;

(13) By adjusting the parameters, determine the constant ε and δ values suitable for the current data according to the degree of conformity between the obtained fitted straight line and the original curve shape.

In the method for measuring the similarity of time series patterns described in step (2), a similarity distance measure function is established for the meta-pattern and the sequence pattern respectively, and the details are as follows:

(21) For the degree of similarity between two meta-patterns, using the weighted distance method, the pattern sequence of the time series X obtained by the PLR algorithm is S _x ={M ₁ , M ₂ , M ₃ ,..., _Mn } , two meta-patterns M _i =(a _i ,Δt _i ), M _j =(a _j ,Δt _j ),(i,j=1,2,3,...,n) are arbitrarily selected in S _x ,

Denoted as: D _wei (Mi, M _j )=β·|a _i -a _j |+(1-β)·|Δt _i -Δt _j |

In the formula, a _i represents the slope of the fitted line, Δt _j represents the time interval adjacent to point _j , Mi represents the meta-pattern, D _wei (Mi, M _j ) is the weighted distance between the meta-patterns Mi and M _j , 0≤ β≤1;

(22) Combining adjacent meta-patterns in two phases to obtain sequential patterns, the similarity measurement of sequential patterns is based on the similarity of meta-patterns, using dynamic time-warped distance method to measure the similarity of sequential patterns;

Time series X=(x ₁ , x ₂ , x ₃ ,..., x _n ) and time series Y=(y ₁ , y ₂ ,..., y _m ), through the KPRA-PLR algorithm for dimensionality reduction, The pattern sequences obtained by linear representation are:

S _x = {M ₁ , M ₂ , M ₃ , ..., M _p },

S _y = {N ₁ , N ₂ , N ₃ , ..., N _t },

where ( _i = ₁ , 2, 3 _{, .}

N _j = (a _j , Δt _j ) is the meta-pattern of S _y ; the calculation formula for establishing the distance matrix between two pattern sequences is as follows:

DM=(a _ij ) _p×t

In the formula, a _ij =D _wei (M _i , N _j ) is expressed as the weighted distance of the i-th meta-pattern Mi of the pattern sequence S _x and the _{j-th meta-pattern N j of} S _y . The value of a _ij reflects the degree of similarity between the meta-patterns _Mi and N _j . When the meta-patterns _Mi and N _j are more similar, _{the value of a ij is} closer to 0 _; The larger the value of a _ij , the larger the value of a ij;

(23) Similar to the dynamic time-curved distance between time series, search for the best curved path W=w ₁ , w ₂ , .. , w _k in the distance matrix DM=(a _ij ) _p×t , and calculate the cumulative distance The minimum curved path, which is called the dynamic time curved distance of the mode sequence S _x and S _y , is abbreviated as D _dtw (S _x , S _y ), and the following formula is obtained:

In the formula, w is to calculate the distance between two meta-patterns.

In the k-nearest neighbor local anomaly detection algorithm described in step (3), the anomaly score value is equal to the k-nearest neighbor distance of the mode, and the specific calculation process is as follows:

Let k be a positive integer, the k-nearest neighbor reachable distance between instance x and instance t is expressed as:

r _{dist(x, t)} = max{k _distance(t) , d(x, t)}

The local reachability density of an instance x is defined as the reciprocal of the average reachable distance of the k nearest neighbors of the instance x, and the local reachability density formula is as follows:

In the data point set D, the LOF(x) of the local outlier coefficients of the instance x can be defined as:

Among them, the abnormal score of each sequence pattern is calculated according to the DTW algorithm, r _{dist(x, t)} is the abnormal score of each sequence pattern, the instance x represents the abnormal score of a sequence pattern, and the local reachability density is calculated according to the formula, lof represents abnormal factor.

Further, for the singularity problem of the DTW method, the improved method based on the spatial dimension and the time dimension, through the method of adaptive weight, does not introduce additional weight coefficients to weight the sum of the spatial dimension (value) and the time dimension (gradient). , and then calculate the minimum distance of the accumulated path.

Further, the singularity problem in the DTW method includes the SDTW algorithm as the time series similarity measurement method, and the specific process is as follows:

Define the feature F _s (x(i)) as:

In the formula, max(|Δx|) represents the maximum gradient of all time points in the time series, and the gradient of a point is represented by the difference of adjacent points, such as x(i)-x(i-1); Max(|Δx|) The function is to constrain the gradient size of each point in the time series to [-1, 1], so as to combine the time dimension with the space dimension in the form of a ratio, and obtain the distance d(i, j) between the elements in the matrix ) is expressed as follows:

d(i,j)=( _Fs (x(i))- _Fs (y(j))) ² .

Beneficial effects: a set of hydrological data abnormal pattern detection scheme of the present invention realizes the KPRA-PLR algorithm to segment the original sequence, uses weighted distance for the similarity measure of the meta pattern, uses the improved DTW algorithm for the similarity measure of the sequence pattern, based on The anomaly pattern detection principle of similarity measurement uses the k-nearest neighbor local anomaly detection algorithm. Compared with the commonly used abnormal pattern detection method—the symbolic SAX method, the method of similarity measurement is used without fixed segmentation and fixed number of pattern categories. The detected abnormal patterns are more accurate, which provides a reference for the detection of abnormal patterns in hydrological data from the perspective of data analysis.

Description of drawings

Fig. 1 is the overall structure schematic diagram of the method in the present invention;

Fig. 2 is KPRA-PLR algorithm fitting straight line schematic diagram in the present invention;

FIG. 3 is a schematic diagram of flow changes in modes 185 to 195 in the embodiment;

FIG. 4 is a schematic diagram of flow changes in modes 282 to 287 in the embodiment;

FIG. 5 is a schematic diagram of flow changes in modes 290 to 302 in the embodiment;

FIG. 6 is a schematic diagram of flow changes in modes 363 to 370 in the embodiment.

Detailed ways

In order to describe the technical solutions disclosed in the present invention in detail, further description is made below with reference to the accompanying drawings and specific embodiments of the description.

A method for detecting abnormal patterns in hydrological data based on similarity measure, as shown in Figure 1, the invented method for detecting abnormal pattern in hydrological data based on similarity measure includes KPRA-PLR algorithm based on linear segment representation of key points, time series There are three modules of pattern similarity measurement method and K-nearest neighbor principle. In this embodiment, the hydrological data is selected to select the hourly flow (unit is m ³ /s) data of the Longmen hydrological station during the flood season, and the station is used from 2000/6/1 6:00:00 to 2015/9/30 17:39:00 The measured hourly flow data in the flood season were used for experiments.

First, the KPRA-PLR algorithm is based on the linear segment representation of key points. According to the formula in the definition of key points, the same thresholds ε and δ are used for the dimensionality reduction of the above two hydrological data, and the historical hydrological time series (2000- 2015 flood season data) and real-time hydrological time series (2016-2017 flood season data) key points, according to statistics: real-time hydrological data obtained 189 key points, historical hydrological data obtained 2844 key points. The key points are represented in the format of [t,i], where t represents the time series index, and i represents the size of the observation value at the corresponding time of the time series. The key points obtained by intercepting part of the real-time hydrological time are displayed in a table, as shown in Table 1 below.

Table 1 shows the key points

[1,159] [4,258] [8,168] [11,244] [12,232] [14,220] [29,188] [32,184] [35,216] [38,183] [43,256] [46,214] [50,205] [52,276] [61,173] [64,406] [65,344] [67,537] [73,330] [75,421] [89,142] [96,164] [105,141] [111,160] [120,387] [121,511], [122,475] [128,615] [131,960] [142,273] [148,768] [149,747] [150,780] [151,705] [155,556] [165,147] [174,396] [177,651] [180,476] [181,554] [184,331] [190,234] [194,1160] [195,1300], [196,1150] [197,125] [198,122] [89,142] … … … … … … [605,296] [606,266] [610,185] [612,181] [615,451] [618,360] [621,580] [624,308] [627,525] [631,388] [635,808] [639,490] [640,514] [643,332] [645,377] [649,232] [651,226] [653,215] [654,266] [657,184] [658,201] [664,480] [666,301] [668,426] [670,301] [672,499] [675,306] [677,322] [678,310] [680,322]

The original time series is divided by key points to obtain sub-sequences, and the sub-sequences are represented by the PLR algorithm to obtain the sub-sequences. According to the definition, adjacent meta-patterns are combined to obtain sequential patterns. If different combination methods are used, the obtained sequential patterns will be different. This chapter uses the combination of adjacent meta-patterns to obtain sequential patterns, using Mi _{= (a i ,} Δt _i ), where a _i represents the slope of the straight line, which represents the trend of the curve, and Δt _i represents the length of the straight line, which represents the length of the curve.

According to statistics: 188 meta-patterns are obtained from the real-time hydrological sequence, and 94 are formed by two adjacent meta-patterns; there are 2,844 meta-patterns in the historical hydrological sequence, and 1,422 are formed by two adjacent meta-patterns. . A partial presentation of the sequence patterns obtained from real-time data is shown in Table 2 below.

Table 2 Partial sequence pattern table

[33.0,3] [-22.5,4] [25.33,3] [-12.0,1] [-6.0,2] [-1.2,5] [-2.6,10] [-1.32,3] [10.67,3] [-11.0,3] [14.6,5] [-14.0,3] [15.0,2] [-19.5,2] [35.5,2] [-11.43,9] [77.67,3] [-62.0,1] [96.5,2] [-43.25,4], [-17.0,2 [45.5,2] [-19.9,14] [3.14,7] [-2.54,9] [3.17,6] [42.25,8] [-111.0,1] [124.0,1] [-36.0,1] [23.33,6] [115.0,3] [-62.44,11] [82.5,6] [-21.0,1] [33.0,1] [-75.0,1] [-37.25,4] [91.4,10] [-119.3,9] [85.0,3] [-58.3,3] [78.0,1] [-74.32,3] [334.83,6] [-295.0,4] [140.0,1] [-150,1] [100.0,1] [-30.0,1] [13.33,3] [75.0,2] [-85.5,6] [32.0,1] … … … … … … [-15.0,4] [20.13,8] [-18.0,6] [18.5,4] [-4.0,4] [-23.3,3] [15.0,4] [24.67,3] [-19.0,6] [13.33,3] [-13.0,1] [27.3,3] [-30.0,1] [-20.25,4] [-2.0,2] [90.0,3] [-30.32,3] [73.3,3] [-90.6,3] [72.33,3] [-34.25,4] [105.0,4] [-79.5,4] [24.0,1] [-60.66,3] [22.5,2] [-36.25,4] [-3.0,2] [-5.5,2], [51.0,1] [99.0,2] [8.0,2] [-12.0,1] [6.0,2]

In the method for measuring the similarity of time series patterns, the experiments in this chapter are mainly to mine abnormal patterns in real-time hydrological data. Therefore, the real-time hydrological data is segmented according to the KPRA-PLR algorithm, and the pattern is represented as a meta-pattern, which consists of two adjacent meta-patterns. Sequence patterns, a total of 94 sequence patterns were obtained. The historical hydrological data were segmented, and the patterns were represented as meta-patterns. Two adjacent meta-patterns formed a sequence pattern, and a total of 1422 sequence patterns were obtained. The distance between each sequence pattern and the 1422 sequence patterns obtained from the historical hydrological data is calculated in pairs, and the results of some similarity measures are shown in Table 3 below.

Table 3 Similarity measurement results

serial number 1 2 3 … 1420 1421 1422 1 9.4 11.91 11.5 … 69.83 3.83 52.15 2 11.91 12.68 16.41 … 75.74 12.08 71.76 3 11.5 16.41 6.35 … 80.33 9.67 25.35 4 25.03 28.12 17.13 …. 93.86 23.2 24.88 5 8.83 8.33 8.08 … 77.66 7.0 6.43 6 50.665 56.575 60.165 … 20.165 53.495 54.815 7 18.67 24.58 29.17 … 53.16 19.5 22.82 8 27.74 30.83 21.24 … 96.57 24.91 27.59 9 26.685 23.985 18.395 … 95.725 25.065 21.745 10 33.895 31.985 26.395 … 102.725 32.065 28.745 11 33.58 35.67 24.08 … 102.41 31.75 32.43 12 12.83 23.74 11.33 … 73.0 13.66 12.98 … … … … … … … … 85 66.58 69.67 59.08 … 135.41 62.75 66.43 86 55.41 58.5 48.91 … 124.24 52.58 55.26 87 30.955 34.045 23.455 … 99.785 28.125 30.805 88 42.08 45.17 36.58 … 110.91 38.25 41.93 89 36.49 39.58 29.99 … 105.32 32.66 37.34 90 58.67 64.58 65.17 … 41.34 60.5 68.82 91 51.17 57.08 61.67 … 19.66 53.0 56.32 92 69.83 75.74 80.33 … 54.9 72.66 74.98 93 3.83 12.08 9.67 … 72.66 56.4 5.68 94 5.15 10.76 11.35 … 73.98 10.68 9.4

In the K-nearest neighbor principle, the SDTW algorithm is used to calculate the similarity between the sequence pattern obtained from real-time hydrological data and the sequence pattern obtained from historical hydrological data. The similarity measurement results are shown in Tables 4 and 3 above. Then calculate the anomaly score of each sequence pattern in the real-time hydrological data. The anomaly score is calculated as: the reciprocal of the average distance between the 1422 sequence patterns obtained from the historical data. After calculating the anomaly score, a similarity-based point anomaly detection algorithm can be used to detect anomalous patterns. There are two methods that can be used, one is the k-nearest neighbor principle; the other is a clustering-based method.

In this chapter, the k-nearest neighbor method is used to calculate the local anomaly factor, and the sequence pattern corresponding to the local anomaly factor with the largest Top-k is selected as the anomaly pattern. An important parameter of the K-nearest neighbor principle is the number of neighbors K. By continuously adjusting the value of the number of neighbors K, the corresponding local anomaly factor is calculated, and the local anomaly factor with the largest Top-k is selected as shown in Table 4 below.

Table 4 Top-k local abnormal factor table

According to the analysis of the experimental results, the local abnormal factors are sorted from large to small, the mode 185-195 is the largest, and the mode 282-287 is the second. From Figure 3 and Figure 4, it can be found that the hourly flow of the current station suddenly surges within 2 hours, and the slope a _i of the sequence pattern is very large, reflecting the rapid growth rate of the sequence pattern, and then decreases. Such a sequence pattern exists in the flow data from 2000 to 2015, but the number of occurrences is very small, and the trend and time length of the sequence pattern are similar. Figure 3 and Figure 4 show that the traffic flow increases too fast in just 2 hours. Secondly, the increase and decrease time of the modes 185 to 195 is too long, and the increase and decrease time of the modes 282 to 287 is too short. Therefore, the mode 185 to 195, and modes 282 to 287 are judged as abnormal modes.

Sort the local anomaly factors from large to small, mode 290-302 is the third, mode 363-370 is the fourth. From Figure 5 and Figure 6, it can be found that the flow increases slowly for a short time, and then decreases for a long time. The slope a _i of the sequence mode is a negative value, reflecting the degree of decline of the sequence mode. The overall time of the sequence mode in Figure 5 is too long , according to the definition of the key point, 1520 is not the key point, there is a downward-rising trend, the first hour is a sharp increase, and the subsequent decline rate quickly drops from 2640 to 755. In the traffic data from 2000 to 2015 There are very few that can be similar to the increasing and decreasing trends and time lengths of patterns 290-302. Figure 6 This sequence pattern only has a slow increase in traffic for a short period of 1 hour, and then it is in a state of decline for a long time. Such a trend seldom occurs, and if the decline time is too long, it appears "strange". Therefore, patterns 290 to 302, pattern 363 ~370 is determined as abnormal mode.

Claims (6)

1. a kind of hydrological data abnormal pattern detection method based on similarity measure, it is characterized in that: described method is based on the piecewise linear representation KPRA-PLR algorithm of key point, the method for time series pattern similarity measure and based on k-nearest neighbor The local anomaly detection algorithm includes the following steps: (1) KPRA-PLR algorithm based on piecewise linear representation of key points: First, the original hydrological time series is divided according to the definition of key points to obtain each subsequence, and then each subsequence is fitted with a straight line according to the PLR algorithm, Obtain the slope of the straight line, solve the intersection between two adjacent straight lines, and obtain the start time t _il and end time t _ir of each subsequence, and use ( _ai , t _il , t _ir ) to represent each subsequence, also known as the element model; (2) Method for measuring the similarity of time series patterns: combine adjacent meta-patterns into sequential patterns; when measuring the similarity of two patterns, establish a similarity distance measure function between the two patterns, which is expressed by the distance The similarity between the two, the average similarity of each pattern represents the abnormal score of the pattern; (3) k-nearest neighbor local anomaly detection algorithm: the k-nearest neighbor distance of the pattern is represented by the anomaly score value, and the k-nearest neighbor local reachable distance, local reachability density, and local anomaly factor are obtained. The largest anomaly factor is the anomaly pattern. . 2. the hydrological data abnormal pattern detection method based on similarity measure according to claim 1, is characterized in that: step (1) comprises following process: (11) Define the key point as S={s ₁ , s ₂ ,...,s _m } as all local extreme points of the time series X, when the extreme point _sk satisfies |s _k -s _k-1 |< When ε and |t _k -t _k-1 |<δ, where ε and δ are given constants, and t _k represents the observation time of each hydrological physical quantity, then s _k is called an unimportant local extreme point. Other extreme points are key points; (12) According to the linear segmental representation algorithm, use the least squares method to fit a straight line through two adjacent key points to obtain the slope corresponding to the straight line, solve the intersection between the two adjacent straight lines, and obtain the starting time and sum of each subsequence. Termination time t _il and t _ir ; the pattern of each subsequence is expressed as (a _i , t _il , t _ir ) is called a meta-pattern; (13) By adjusting the parameters, determine the constant ε and δ values suitable for the current data according to the degree of conformity between the obtained fitted straight line and the original curve shape. 3. The method for detecting abnormal patterns of hydrological data based on similarity measurement according to claim 1, wherein: in the method for time series pattern similarity measurement described in step (2), the meta-pattern and the sequence pattern are established respectively The similarity distance metric function is as follows: (21) For the degree of similarity between two meta-patterns, using the weighted distance method, the pattern sequence of the time series X obtained by the PLR algorithm is S _x ={M ₁ ,M ₂ ,M ₃ ,...,M _n }, S Two meta-patterns M _i =(a _i ,Δt _i ), M _j =(a _j ,Δt _j ),(i,j=1,2,3,...,n) are arbitrarily selected in _x , Denoted as: D _wei (M _i ,M _j )=β·|a _i -a _j |+(1-β)·|Δt _i -Δt _j | In the formula, a _i represents the slope of the fitted line, Δt _j represents the time interval adjacent to point _j , Mi represents the meta-pattern, D _wei (Mi, M _j ) is the weighted distance between the meta-patterns Mi and M _j , 0≤ β≤1; (22) Combining adjacent meta-patterns in two phases to obtain sequential patterns, the similarity measurement of sequential patterns is based on the similarity of meta-patterns, using dynamic time-warped distance method to measure the similarity of sequential patterns; Time series X=(x ₁ , x ₂ , x ₃ ,...,x _n ) and time series Y=(y ₁ , y ₂ ,..., y _m ), obtained by dimensionality reduction and linear representation by KPRA-PLR algorithm The pattern sequences are: S _x ={M ₁ ,M ₂ ,M ₃ ,...,M _p }, S _y ={N ₁ ,N ₂ ,N ₃ ,...,N _t }, where (i=1,2,3,...,p; j=1,2,3,...,t), M _i =(a _i ,Δt _i ) is the meta-pattern of S _x , N _j =(a _j ,Δt _j ) is the meta-pattern of S _y ; the calculation formula for establishing the distance matrix between two pattern sequences is as follows: DM=(a _ij ) _p×t In the formula, a _ij =D _wei (M _i , N _j ) represents the weighted distance between the i-th meta-pattern Mi of the pattern sequence S _x and the _{j-th meta-pattern N j of} S _y . The value of a _ij reflects the degree of similarity between the meta-patterns _Mi and N _j . When the meta-patterns _Mi and N _j are more similar, _{the value of a ij is} closer to 0 _; The larger the value of a _ij , the larger the value of a ij; (23) Similar to the dynamic time warp distance between time series, search for the best curved path W=w ₁ ,w ₂ ,...,w _k in the distance matrix DM=(a _ij ) _p×t , and calculate the minimum cumulative distance The curved path of , this minimum value is referred to as the dynamic time curved distance of the mode sequence S _x and S _y , abbreviated as D _dtw (S _x , S _y ), and the following formula is obtained: In the formula, w is to calculate the distance between two meta-patterns. 4. the hydrological data abnormal pattern detection method based on similarity measure according to claim 1 is characterized in that: in the described k-nearest neighbor local abnormality detection algorithm of step (3), the abnormal score value is equal to the k- of this pattern The nearest neighbor distance, the specific process is as follows: Let k be a positive integer, the k-nearest neighbor reachable distance between instance x and instance t is expressed as: r _dist(x,t) =max{k _distance(t) ,d(x,t)} The local reachability density of an instance x is defined as the inverse of the average reachable distance of the k nearest neighbors of the instance x, and the local reachability density formula is as follows: In the data point set D, the LOF(x) of the local outlier coefficients of the instance x can be defined as: Among them, the anomaly score of each sequence pattern is calculated according to the DTW algorithm, r _dist(x,t) is the anomaly score of each sequence pattern, the instance x represents the anomaly score of a sequence pattern, and the local reachability density is calculated according to the formula, lof represents abnormal factor. 5. the hydrological data abnormal pattern detection method based on similarity measure according to claim 3, it is characterized in that: for the singular point problem that DTW method exists, the improvement method based on space dimension, time dimension, by the mode of self-adaptive weight , without introducing additional weight coefficients to weight the space dimension and time dimension, and then calculate the minimum distance of the accumulated path. 6. the abnormal pattern detection method of hydrological data based on similarity measure according to claim 5, is characterized in that: the singularity problem in described DTW method comprises to carry out data processing as time series similarity measure method based on SDTW algorithm, The specific process is as follows: Define the feature F _s (x(i)) as: where max(|Δx|) represents the maximum gradient of all time points in the time series, and the gradient of a point is represented by the difference between adjacent points, such as x(i)-x(i-1); Max(|Δx|) The function is to constrain the gradient size of each point in the time series to [-1, 1], so as to combine the time dimension with the space dimension in the form of a ratio, and obtain the distance d(i, j) between the elements in the matrix ) is expressed as follows: d(i,j)=(F _s (x(i))-F _s (y(j))) ² .