JP2013257712A - Device, method and program for detection of similar partial sequence - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately detect a similar partial sequence even with a small amount of data without using all of time series data.SOLUTION: When one element x of sampled time series data is received (S201), a similarity score between a partial sequence of time series data including the element, and a partial sequence of another time series data, is calculated using a sampling period (S103). A pair of partial sequences for which the similarity score is equal to or more than a predetermined threshold is determined as a pair of similar partial sequences, and the pair of similar partial sequences is detected as a pair of matching partial sequences (S104).

Description

  The present invention relates to a similar partial sequence detection device, method, and program, and more particularly, to a similar partial sequence detection device, method, and program for detecting a pair of similar partial sequences from two time-series data.

  With the widespread use of sensor networks, a technique for efficiently processing and analyzing sensor data collected from sensor nodes has come to be demanded. In particular, sensor data is not analyzed after accumulating data, as represented by recent big data, but it is important to perform analysis in real time and present the results. In addition to statistical information processing such as obtaining the average value and maximum / minimum values of sensor data that have been studied in the past, advanced mining technology that automatically extracts useful information is required. One of the mining techniques is similarity search. Similarity search of sensor data is a technique for specifying a similar part between sensor data, such as correlation detection (for example, Patent Document 1) and identification of a partial pattern similar to an inquiry pattern (for example, Patent Document 2). It corresponds to. Patent Document 3 discloses a technique for specifying a partial pattern common between data. The dynamic time warping distance (DTW) is used as a distance function for pattern detection, and it operates at high speed and saves memory while ensuring the same accuracy as when using pure DTW. This means that sensor data is monitored by a server or the like, and if there is a common pattern between two sensors, it is detected in real time. FIG. 17 and FIG. 18 illustrate specific similar pattern detection. FIG. 17 is humidity data acquired by two different sensors. Each sensor acquires data at intervals of about 1 minute, the horizontal axis is the sequence number of the acquired data, and the vertical axis is the value. These two sensor data show very similar patterns, although there are some different sections. Specifically, the two patterns described by Subseq are similar patterns in these sensors. FIG. 18 shows similar pattern detection results when the technique of Patent Document 3 is applied to these data. The horizontal axis represents the sequence number of the first sensor data in FIG. 17, and the vertical axis represents the sequence number of the second sensor data. The bold lines in the figure indicate which elements correspond to each other between the two data, and show how two similar patterns are specified. In this way, the two sensor data are compared, and if there is a similar pattern, the position (start position and end position) and similarity are reported in real time. This technique is a general-purpose technique applied to time series data, but when applied to sensor data, it is necessary to pay attention to the following two points. The first point is the amount of data communication. In order to detect a similar pattern, it is necessary to transmit data acquired by each sensor node to a server. At this time, if the data is acquired at short intervals, the amount of data becomes enormous. Since each sensor node transmits all of these, the communication amount of the entire network increases. The second point is the power restriction of the sensor node. Each sensor node is mainly battery-powered and consumes power during data collection and data transmission. This amount of power is known to be the highest during data transmission, and transmitting more data than necessary may accelerate battery consumption. From the above, when similar pattern detection is performed in real time from sensor data, it is necessary to efficiently transmit data while retaining minimum data necessary for pattern detection.

  Therefore, as an efficient data transmission method, data is selected and transmitted according to the rate of change of data (calculated based on speed, position, etc.) as in Patent Document 4, or plural data as in Patent Document 5 A method of compressing and transmitting the sensor data is available.

  Further, Non-Patent Document 1 proposes a technique for converting, in real time, only time-series data into summary information obtained by reducing original data. By transmitting a part of the summary information, the amount of data is greatly reduced.

Japanese Patent No.4452234 Japanese Patent No. 4801566 JP 2010-198227 JP 2011-192198A JP 2012-44449 A

Umit Y. Ogras, and Hakan Ferhatosmanoglu: “Online summarizationof dynamic time series data,” VLDB journal, vol.15, num.1, pp.84-98, 2006.

  The DTW used in Patent Document 3 described above detects a similar pattern using the shape of data (time domain information). Therefore, in order to guarantee the detection accuracy close to the original data, it is necessary to select the data so that the shape of the data is maintained while realizing the data reduction. On the other hand, in the technique described above, data is selected regardless of the data shape or in a state where the data shape is broken, so that it is difficult to appropriately detect similar patterns.

  The present invention has been made in view of the above circumstances, and can detect a similar partial sequence with high accuracy even with a small amount of data without using all data (original data) of time-series data. An object is to provide a similar partial sequence detection apparatus, method, and program.

  In order to achieve the above object, a similar partial sequence detection apparatus according to the present invention samples first time-series data obtained by sampling first time-series data at a first sampling period and second time-series data at a second sampling period. A single score matrix having a similarity score that can be converted into a DTW (Dynamic Time Warping) distance between the two partial sequences as a matrix element from the second time series data, A similar partial sequence detection device that detects using a position matrix having a start position of the partial sequence as a matrix element, the storage unit storing the score matrix and the position matrix, the sampled first time series data, When one element of any data of the second time series data is received, the first time series data including the element and The similarity score between the partial sequence in any one of the two time series data and the partial sequence in any one of the first time series data and the second time series data is calculated, and the calculated similarity score is The two partial sequences used for calculating the similarity score are stored as matrix elements of the score matrix of the storage unit corresponding to the end positions of the two partial sequences used for the calculation of the similarity score. A pair of partial sequences in which the start position is stored as a matrix element of the position matrix of the storage unit corresponding to the end position, and the similarity score stored in the score matrix of the storage unit is greater than or equal to a predetermined threshold A processing unit for determining a pair of similar partial sequences and detecting the pair of similar partial sequences as a pair of matching partial sequences, The processing unit obtains the similarity score adjacent to the similarity score in the score matrix and corresponding to the previous time with respect to the first time-series data when calculating any of the similarity scores, A value obtained by multiplying the obtained similarity score by the first sampling period is subtracted from the magnitude of the difference between corresponding data elements in the two partial sequences of interest, and a predetermined coefficient is obtained. A value obtained by multiplying a numerical value corresponding to the first sampling period is added to calculate a first score, which is adjacent to the similarity score in the score matrix and corresponds to the previous time with respect to the second time-series data. The similarity score is acquired, and with respect to the acquired similarity score, the magnitude of a difference between corresponding data elements in the two target partial sequences is determined. Subtracting the value multiplied by the second sampling period and adding a value obtained by multiplying the predetermined coefficient by a numerical value corresponding to the second sampling period to calculate a second score; The similarity score that is adjacent to the similarity score and corresponds to the previous time for both the first time series data and the second time series data is acquired, and for the acquired similarity score, A value obtained by multiplying a magnitude of a difference between corresponding data elements in the two partial sequences by a larger one of the first sampling period and the second sampling frequency is subtracted, and the predetermined coefficient is A value obtained by multiplying a numerical value corresponding to the sum of one sampling period and the second sampling period is added to calculate a third score, and the first score, the second score are calculated. A, and a maximum score among the third score, and the similarity score.

  The similar partial sequence detection method according to the present invention is based on the first time series data obtained by sampling the first time series data in the first sampling period and the second time series data obtained by sampling the second time series data in the second sampling period. A single score matrix having a similarity score that can be interconverted between a pair of similar partial sequences as a DTW (Dynamic Time Warping) distance between the two partial sequences, and a start position of the partial sequence as a matrix A similar partial sequence detection method in a similar partial sequence detection apparatus comprising a storage unit for storing the score matrix and the position matrix, which is detected using a position matrix as an element, wherein the sampled first sampled by a processing unit When one element of either time-series data or second time-series data is received, A similarity score between a partial sequence in any one of the first time-series data and the second time-series data including the prime, and a partial sequence in any one of the first time-series data and the second time-series data, Calculating and storing the calculated similarity score as a matrix element of the score matrix of the storage unit corresponding to the end positions of the two partial sequences used to calculate the similarity score, and the similarity score Storing the start positions of the two partial sequences used for the calculation as matrix elements of the position matrix of the storage unit corresponding to the end positions, and the similarity score stored in the score matrix of the storage unit is A partial sequence pair that is equal to or greater than a predetermined threshold is determined as a similar partial sequence pair, and the similar partial sequence pair is determined as a matching partial sequence pair. And detecting the similarity score by the processing unit, and the similarity corresponding to the previous time with respect to the first time series data adjacent to the similarity score in the score matrix The degree score is acquired, and the value obtained by multiplying the magnitude of the difference between corresponding data elements in the two partial sequences of interest by the first sampling period is subtracted from the acquired similarity score, A value obtained by multiplying a predetermined coefficient by a numerical value corresponding to the first sampling period is added to calculate a first score, and one of the second time-series data adjacent to the similarity score in the score matrix is calculated. The similarity score corresponding to the previous time is acquired, and the corresponding similarity score in the two partial sequences of the target is acquired with respect to the acquired similarity score. Subtracting the value obtained by multiplying the difference between the data elements by the second sampling period, and adding the value obtained by multiplying the predetermined coefficient by a numerical value corresponding to the second sampling period, 2 scores are calculated, and the similarity score adjacent to the similarity score in the score matrix and corresponding to the previous time for both the first time series data and the second time series data is obtained, A value obtained by multiplying the obtained similarity score by the larger of the first sampling period and the second sampling frequency to the magnitude of the difference between corresponding data elements in the two target partial sequences is subtracted. And a third score is calculated by adding a value obtained by multiplying the predetermined coefficient by a numerical value corresponding to the sum of the first sampling period and the second sampling period. The first score and the second score, and the highest score among the third score, and the similarity score.

  The program according to the present invention is a program for causing a computer to function as each unit of the similar partial sequence detection device.

  As described above, according to the similar partial sequence detection device, method, and program of the present invention, by calculating the similarity score using the sampling period of each time series data, all of the time series data is calculated. An effect is obtained that a similar partial sequence can be detected with high accuracy even when there is little data without using data.

It is a graph which shows the power spectrum of humidity data. It is a graph which shows the humidity data after sampling. It is the schematic which shows the structure of the sensor node which concerns on the 1st Embodiment of this invention. It is the schematic which shows the structure of the similar pattern detection apparatus which concerns on the 1st Embodiment of this invention. It is a figure for demonstrating DTW. It is a flowchart which shows the content of the processing routine of the similar partial sequence pair detection process in the similar pattern detection apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the processing routine of the storing process to List arrangement | sequence in the similar pattern detection apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the processing routine of the storing process to List arrangement | sequence in the similar pattern detection apparatus which concerns on the 1st Embodiment of this invention. It is the schematic which shows the structure of the sensor node which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the content of the processing routine of the sampling period determination process in the sensor node which concerns on the 2nd Embodiment of this invention. It is a graph which shows the data of a temperature sensor. It is a figure which shows the similar pattern detection result using the data of a temperature sensor. It is a graph which shows the data of a sunspot. It is a figure which shows the similar pattern detection result using the data of a sunspot. It is a figure which shows the similar pattern detection result using the data of a temperature sensor. It is a figure which shows the similar pattern detection result using the data of a sunspot. It is a figure for demonstrating the partial pattern which is common with humidity data. It is a figure which shows the example of a detection of the similar pattern using a prior art.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
<Outline of the invention>
The present invention provides a system that maintains an accuracy close to that of the conventional method (Patent Document 3) while realizing efficient data transmission to a sensor network in detecting a similar pattern between sensor data. That is, an object of the present invention is to detect a similar pattern with high accuracy even with a small amount of data without using all data (original data) collected by the sensor node. The sensor node does not transmit all collected data, but selects and transmits only data necessary for similar pattern detection. On the other hand, the server detects from the transmitted data while maintaining the accuracy close to the similar pattern detection using the original data.

In order to achieve the same accuracy as the original data, it is necessary to select (sample) the data so as to maintain the shape of the data. Therefore, in the present invention, similar pattern detection is performed by sampling original data based on the sampling theorem and transmitting the sampled data. The sampling theorem is a theorem that quantitatively indicates how much sampling should be performed when converting an analog signal to a digital signal, where f is the maximum frequency component contained in the original signal, A signal sampled at a frequency f _Nq higher than 2f means that the original signal can be completely decoded. This f _Nq is called the Nyquist frequency. In order to determine the sampling period, discrete Fourier transform (FFT) is performed on the data to calculate a power spectrum. This represents how much each frequency component is included in the data. FIG. 1 illustrates the power spectrum of Humidity # 1. The power in FIG. 1 is normalized so that the sum is 1. Although FIG. 1 shows up to 60 frequency components, it can be seen that power is also allocated to higher frequency components, which are composed of various frequency components. In order to determine the maximum frequency component f of such data, a high frequency component with a low power ratio is considered to be noise and is generally determined using a threshold value (for example, signal processing or network analysis). ). There are two main determination methods based on the threshold, the ratio to the total power is specified as the threshold, the frequency where the total power from the low frequency components exceeds the threshold is set as f, and the power value is specified as the threshold. Of the frequency components having a power equal to or higher than a threshold value, the maximum frequency component may be defined as f. However, the method for determining the maximum frequency component is not limited to this, and any method may be used. Next, based on f, the Nyquist frequency f _Nq is obtained by the following equation (1), and the sampling period T is determined according to the following equation (2).

  Here, n represents the length of data. The sampling period may be used in common for all sensor nodes, or may be set individually for each sensor node. However, since the sampling period is a value that depends on the data, it is desirable to set it individually.

  Each sensor node samples the original data according to the set sampling cycle. Then, only the sampled data is transmitted to the similar pattern detection apparatus which is a server. FIG. 2 shows Humidity # 1 data after sampling. Despite maintaining the shape of the data, its data size has been greatly reduced.

  On the other hand, information on the sampling period is used when detecting a similar pattern in the similar pattern detection apparatus. That is, instead of using a sampled sequence purely, information on data that has not been transmitted by sampling is approximately used to calculate the similarity. As a result, it is possible to obtain a result close to that when the original sequence is used.

<System configuration>
As shown in FIG. 3, the sensor node 100 according to the first embodiment of the present invention includes a data sensing unit 10, a data acquisition unit 12, a sampling unit 14, a period storage unit 16, and a wireless communication unit 18. The Each data is detected by the data sensing unit 10 and sent to the data acquisition unit 12. The cycle storage unit 16 stores information (sampling cycle) for sampling data, and the sampling unit 14 samples data based on the information. The sensor node 100 obtains information on a sampling cycle for selecting data from a similar pattern detection device 150 described later.

  The time series data sampled by the sampling unit 14 is transmitted from the wireless communication unit 18 to the similar pattern detection device 150.

  As shown in FIG. 4, the similar pattern detection apparatus 150 receives sampled time-series data transmitted from two sensor nodes 100 and detects similar partial sequences. The similar pattern detection device 150 is configured by a server including a CPU, a RAM, and a ROM that stores a program for executing various processing routines described later, and is functionally configured as follows. Yes. As shown in FIG. 4, the similar pattern detection apparatus 150 includes a data input unit 60, a processing unit 70, and a data output unit 80.

The data input unit 60 receives input of a plurality of time series data received by the wireless communication device. In addition, the data input unit 60 receives an input of a similar partial sequence detection condition for detecting a similar partial sequence pair (hereinafter also simply referred to as “similar partial sequence”) from an input device such as a keyboard. The similar partial sequence detection condition is, for example, detecting a pair having a lower limit value l _min (preliminarily set by the user. That is, L (l _x , l _y ) of l _min or more. ), A predetermined coefficient ε set in advance according to the type of time-series data. The data input unit 60 may accept data input from the outside via a network or the like.

  The processing unit 70 includes a data acquisition unit 72, a detection condition storage unit 74, a pattern detection unit 76, a cycle calculation unit 78, a cycle storage unit 82, and a matrix information storage unit 84.

  The data acquisition unit 72 acquires time-series data received by the data input unit 60. The detection condition storage unit 74 stores the similar partial sequence detection condition received by the data input unit 60.

  The period calculation unit 78 performs discrete Fourier transform (FFT) on each of the two time series data acquired from the two sensor nodes 100, calculates a power spectrum, determines the frequency f, and (1 ) And (2) respectively, the sampling period is calculated. The sampling cycles calculated for the two sensor nodes 100 are stored in the cycle storage unit 82, and the data output unit 80 sets the calculated sampling cycle to each of the two sensor nodes 100 by wireless communication. Send.

  The pattern detection unit 76 detects similar partial sequences for the two time-series data acquired from the two sensor nodes 100. Calculation data (score matrix, position matrix, List array) necessary for detecting similar partial sequences is stored in the matrix information storage unit 84. Finally, only the similar partial sequence corresponding to the similar partial sequence detection condition is output by the data output unit 80.

  Next, the principle of detecting a similar partial sequence will be described.

  The present invention is for finding partial similarity patterns between sequences of sensor data. The dynamic time warping distance (DTW) is used for similarity determination between sequences. DTW is one of the useful distance measures considering the time scale. Since the sequence length is adjusted in the time axis direction so as to minimize the distance between sequences, the similarity between sequences having different sequence lengths and sampling rates can be calculated.

  The DTW distance between the two sequences is calculated by dynamic programming and is the sum of the distances after optimal adjustment of the two sequences. 5A to 5C illustrate a DTW calculation method. FIG. 5A shows an optimum sequence association by DTW for distance calculation. For the distance calculation, a time warping matrix as shown in FIG. 5B is used. The arrangement of elements associated with each other for calculating the distance value is called a warping path, and is shown as a set of darkly colored cells in FIG.

Consider a sequence X = (x ₁ , x ₂ , ..., x _n ) of length _n and a sequence Y = (y ₁ , y ₂ , ..., y _m ) of length m. These DTW distances D (X, Y) are defined as the following equation (3).

Here, || x _i −y _j || = (x _i −y _j ) ² represents a distance between two numerical values. Other selections such as L1 distance (|| x _i −y _j || = | x _i −y _j |) may be used. FIG. 5C illustrates a specific DTW distance calculation. In each cell, the distance is calculated according to the above equation (3), and the DTW distance between the two sequences is acquired by the upper right cell. A warping constraint is introduced in order to avoid a reduction in matching accuracy due to a correspondence between a small range of one sequence and a large range of the other sequence. This removes unnecessary warping paths and limits the scope of warping, and the Sakoe-Chiba band is one of the typical warping constraints. The cell colored in gray in FIG. 5B represents a warping range w in Sakoe-Chibaband, and is specifically calculated based on a restriction of | i−j | ≦ w. This is also useful for reducing the amount of calculation in detecting similar patterns. Since the sensor data is data transmitted continuously from each sensor node, it can be said that the sensor data is semi-infinite stream data. In such data analysis, new data received at present is more important than old data received in the past. Therefore, by setting the warping range, it is possible to focus on detecting a similar pattern of relatively new data. Furthermore, the amount of calculation is also kept constant for one matrix, leading to efficient processing. Also in the embodiment of the present invention, an example in which a warping range is set is shown.

The sensor data X is a semi-infinite data sequence composed of values of x ₁ , x ₂ ,. x _n is the latest value, and n increases with time. X [i _s : i _e ] is a partial sequence (pattern) of X starting from time is and ending at i _e , and the length of X [i _s : i _e ] is l _x = i _e −i _s +1 Become. Similarly, Y [j _s : j _e ] is a partial sequence of Y that starts at time j _s and ends at j _e , and its length is l _y = j _e −j _s +1.

  The partial similar pattern between sensor data sequences detected by the present invention is to detect a partial sequence pair that satisfies the condition shown in the following equation (4).

D (X [i _s : i _e ], Y [j _s : j _e ]) is the DTW distance between the partial sequence pair X [i _s : i _e ] and Y [j _s : j _e ], and L is It is a function representing the length of a partial sequence pair. Hereinafter, the partial sequence that satisfies the condition of the above expression (4) is also referred to as a suitable partial sequence. In the present invention, L (l _x , l _y ) = (l _x + l _y ) = 2, which is the average length of two partial sequences, is used, but L (l _x , l _y ) = max (l _x , l _y ) or L (l _x , l _y ) = min (l _x , l _y ) may be used. ε represents a threshold value for similarity determination. Since the DTW distance is represented by the sum of the distances of the associated elements, the value increases as the length of the partial sequence increases. Therefore, it is desirable that the threshold for similarity determination is also proportional to the partial sequence length, and determination is performed using ε (L (l _x , l _y ) −l _min ). The threshold value l _min of the length of the matching partial sequence is a value given by the user, and it is guaranteed that a partial sequence pair whose length L of the partial sequence pair is equal to or greater than l _min is detected.

When the partial sequence pairs X [i _s : i _e ] and Y [j _s : j _e ] are matched, many partial sequence pairs that overlap the partial sequence pair whose distance value takes the minimum value are also matched. End up. Overlap means that warping paths between partial sequence pairs intersect, and that at least one element is shared in two warping paths. The overlapping partial sequence pairs are redundant information, and a plurality of similar patterns are detected from substantially the same section. Also, reporting unnecessary results may reduce the processing speed of the algorithm. Therefore, a partial sequence pair having a minimum distance value is detected from among the overlapping partial sequence pairs. Specifically, a partial sequence pair that satisfies the following conditions (1) and (2) is detected as a similar pattern.

(Problem 1) Given two sequences X and Y, a threshold value ε, and a threshold value l _min for the length of a matching partial sequence, a partial sequence pair X [i that satisfies the following conditions (1) and (2) _s : i _e ] and Y [j _s : j _e ] are detected.

(1) X [i _s : i _e ] and Y [j _s : j _e ] satisfy the above equation (4).
(2) Among the groups of overlapping partial sequence pairs, D (X [i _s : i _e ]; Y [j _s : j _e ]) − ε (L (l _x ; l _y ) −l _min ) Take the minimum value.

In the first embodiment, each sensor node 100 transmits data based on a fixed sampling period, and similar pattern detection is performed using the reduced (sampled) data. The sampling period of each sensor node 100 is determined on the similar pattern detection device 150. Since the similar pattern detection apparatus 150 determines the sampling period, it is necessary to receive all data until the sampling period is determined. Thereafter, each sensor node 100 transmits the sampled data to the similar pattern detection device 150 based on the determined sampling period. Similar pattern detection can also be performed on the sampled data using the technique of Patent Document 3 as it is. However, since it is different from the original data collected by the sensor node 100, there is a possibility that the result is greatly different from the similar pattern detection using the original data. Therefore, in order to obtain a detection result closer to the original data, the detection method in Patent Document 3 is improved. Specifically, pattern detection is performed by approximately incorporating data information that has not been transmitted by sampling. As a result, it is possible to achieve an accuracy close to that of the original data while realizing an efficient data transmission. In the above Patent Document 3, in order to realize similar pattern detection based on DTW, similarity is calculated using a scoring function and a position matrix that achieve accuracy equivalent to that of DTW. The scoring function is a function that outputs the similarity of a partial sequence pair between two sequences as a score. The score is calculated using a score matrix based on dynamic programming. Specifically, two sequences X = (x ₁ , ..., x _i , ..., x _n ) and Y = (y ₁ , ..., y _j , ..., ..., when y _m) is _{given, X [i s: i e} ] and Y [j _s: j _e] score value V (X a _{[i s: i e];} Y [j s: j e]) is And is calculated as in the following equation (5).

An arrangement of elements associated with each other to calculate a score value is a warping path in the score matrix. Since the score value is determined by accumulating the difference between the distance threshold ε and the distance value of each element, the score value indicates a negative value when the partial sequence pair does not satisfy the condition. Therefore, it becomes possible to determine a non-conforming partial sequence pair from the score value. In addition, by resetting the score of an element having a negative value to zero, score calculation is newly started from that element. By utilizing this property, it is possible to prun out incompatible partial sequence pairs and efficiently calculate only partial sequence pairs that are highly likely to be matched with one matrix. In the above equation (5), b _v , b _h , and b _d are weights determined by L. For example, when L (l _x , l _y ) = (l _x + l _y ) = 2, b _v = b _h = 1/2 and b _d = 1. This is because the sequence length increases by 1/2 when a vertical or horizontal element is inherited and increases by 1 when a diagonal element is inherited. In addition, in L (l _x , l _y ) = max (l _x , l _y ), when l _x > l _y , b _d = b _h = 1, b _v = 0, l _x <l _y , When b _d = b _v = 1, b _h = 0 and l _x = l _y , b _d = 1, b _v = b _h = 0. L (l _x , l _y ) = min (l _x , l _y ) is similarly determined.

The sum of the weights (b _v , b _h , b _d ) on each warping path in the score matrix is designed to be equal to L (l _x ; l _y ), and the DTW distance and score are converted It is guaranteed that it is possible. The DTW distance of the partial sequence pair is calculated by the following equation (6).

  In the score matrix, information on the scores and end points of similar partial sequence pairs is held. Information of the starting point indicating where the partial sequence pairs starting from similar are similar is held by the position matrix. The start point is expressed as a coordinate value, and is obtained as in the following equation (7).

s (i _e , j _e ) indicates the starting point (i _s , j _s ) of matching between X [i _s : i _e ] and Y [j _s : j _e ]. Similar patterns are efficiently detected by using two score matrices.

In the present invention, a similar pattern is detected using a sampled sequence. In order to achieve accuracy close to that of the original data, the scoring function and the position matrix are improved. The intuitive idea is to complement the distance values given by the elements deleted from the original sequence when selecting either vertical, horizontal or diagonal elements in the score calculation. If the sampling period is T, when calculating the score between the sampled sequences, there are T−1 hidden cells between the current cell and the adjacent cells. The distance value in this hidden cell is approximated using the current cell distance value. Sampling period of sequence x of length n and sequence Y of length m is T _x and T _y , and sampled sequence of original sequence X and Y is X '= (x ₁ , ..., x _i , ... ., x _{| _n / Tx "} ), Y '= (y ₁ , ..., y _j , ..., y _{| _m / Ty"} ), the score of the partial sequence of X' and Y 'is And is calculated as in the following equation (8).

  However, | _M ”represents a value obtained by truncating the decimal point of M.

To minimize detection errors, complementing the T _x or T _y-number of distance values when you select a score for each direction. In the above equation (8), the weights of b _v , b _h and b _d are also updated. When L (l _x , l _y ) = (l _x + l _y ) = 2, if a vertical or horizontal cell score is selected, the value of L increases by T _x / 2 or T _y / 2 To do. Similarly, if a diagonal cell score is selected, the value of L increases by (T _x + T _y ) / 2. Therefore, b _v = T _y / 2, b _h = T _x / 2, and b _d = (T _x + T _y ) / 2.

The position matrix is also improved in accordance with the improvement of the above equation (8). However, the operation is almost the same as the above equation (7). If any of the three cells (vertical, horizontal and diagonal cells) is selected, the starting point of that cell is taken over. Otherwise, the starting point is set according to the sampling periods T _x and T _y . That is, the starting point s (i, j) is determined as in the following equation (9).

Next, how to detect similar partial sequence pairs between sampled sequences using a scoring function, a score matrix, and a position matrix will be described. When data is received from a sensor node that is a pattern detection target, detection processing is started. Here, the original data sequences of the two sensor nodes 100 are X and Y (the latest elements are n and m, respectively), and the sampled sequences are X ′ = (x ₁ , ..., x _i , ... , x _{Ln / Tx "), Y '= (y 1} , ..., y j, ..., and y _{Lm / Ty").} When the latest data is received from any one of the sensor nodes 100, the elements of the score matrix and the position matrix corresponding to the data are updated. Then, it checks whether there is a partial sequence pair indicating a score that satisfies the condition (condition 1 of problem 1), and if it exists, checks whether the pair is an optimal pair (problem 1 Condition (2)). Use the List array to find the best pair. When the List array is empty, the score v _max , the start point (i _s , j _s ), and the end point (j _s , j _e ) of similar partial sequence pairs that satisfy the condition are retained. If the List array is not empty, a check is made to see if the newly detected partial sequence pair overlaps with the partial sequence pair stored in the List array. Information of only the partial sequence pair is held in the List array. If there is no overlapping partial sequence pair, it is newly added to the List array. Then, after confirming that the present optimum partial sequence pair is not replaced by a partial sequence pair that appears in the future, that is, when the condition shown in the following equation (10) is satisfied, it is reported as a similar partial sequence pair.

  When each element of the score matrix and the position matrix corresponding to the latest received data is calculated, it is checked whether the partial sequence pair in the List array satisfies the condition shown in the above equation (10). When the calculation of each element of the score matrix and the position matrix corresponding to the latest received data is completed, the partial sequence pair that satisfies the condition of the above expression (10) is reported and then deleted from the List array.

<Operation of Similar Pattern Detection Device>
Next, specific processing of the similar pattern detection apparatus 150 will be described. Here, two time series data X ′ = (x ₁ , ..., x _i , ..., x _{Ln / Tx ”} ) and time series data Y ′ = (y ₁ , ..., y _j , ..., y _{Lm / Ty "} )) and the respective elements are given sequentially. At this time, the similar partial sequence pair detection processing routine shown in FIG.

First, in step S <b _> 101, the data acquisition unit 72 receives the latest data x _{Ln / Tx ”} via the data input unit 60. Next, in step S102, the processing unit 70 resets the value of j to “| _ (m−w) / Ty”.

  Thereafter, the pattern detection unit 76 repeats the processing of steps S103 to S105 until the value of j becomes | _ (m−w) / Ty ”while incrementing the value of j (increasing“ 1 ”).

In step S103, the pattern detection unit 76 uses the similarity score v (n / T _x , j) of x _{n / Tx} and y _j , and the starting point s (n / T _x , j) as the expression (8), ( 9) Calculate using the equation.

In step S104, the pattern detection unit 76 determines whether or not the condition of v (n / T _x , j) −εl _min ≧ 0 is satisfied, and when this condition is satisfied (Yes), the process proceeds to step S105. If the condition is not satisfied (No), the process returns to step S103 or proceeds to step S111.

  In step S105, List array storage processing is performed. The List array storage processing is realized by the processing routine shown in FIG.

In step S120, the pattern detection unit 76 determines whether the List array stored in the matrix information storage unit 84 is empty. If the List array is empty, in step S121, the pattern detection unit 76 stores the score v (n / T _x , j) in List [1] .v _max of the List array, and ( S (n / T _x , j) is stored in List [1] .i _s , List [1] .j _s ), and (List [1] .i _e , List [1] .j _e in the List array) ), (N, j · T _y ) is stored. Then, the processing routine ends.

  On the other hand, when the List array is not empty, the pattern detection unit 76 repeats the process of Step S122 for each element l (l = 1,..., L) in the List array.

In step S122, the pattern detection unit 76 determines whether or not the start point s (n / T _x , j) matches (List [l] .i _s , List [l] .j _s ) in the List array. If it is determined that they do not match, step S122 is repeated or the process proceeds to step S123.

  In step S123, the pattern detection unit 76 determines whether or not it overlaps with all elements of the List array. As a result of the determination in step S122, if all elements do not match the start point, it is determined that there is no overlap, and in step S124,

The score v (n / T _x , j) is stored in List [L + 1] .v _max of the List array, and s (is stored in (List [L + 1] .i _s , List [L + 1] .j _s ) of the List array. n / _{T x,} j) stores, in the List sequence (List [L + 1] .i e, the _{List [L + 1] .j e} ), store the _{(n, j · T y)} , it terminates the processing routine To do.

On the other hand, when it is determined in step S122 that the starting point s (n / T _x , j) matches (List [l] .i _s , List [l] .j _s ) in the List array, the List array In step S125, the pattern detection unit 76 determines that the element overlaps the element.

It is determined whether the score v (n / T _x , j) is equal to or higher than the score List [l] .v _{max of the} element in the List array. If the score v (n / T _x , j) is equal to or higher than the score List [l] .v _{max of the} element of the List array, in step S126, the pattern detection unit 76 lists List [l] of the List array. .v _max is overwritten with a score v (n / T _x , j), and (List [l] .i _s , List [l] .j _s ) of the List array is replaced with s (n / T _x , j) Is overwritten, (n, j · T _y ) is overwritten on (List [l] .i _e , List [l] .j _e ) of the List array, and the processing routine is terminated. On the other hand, when the score v (n / T _x , j) is less than the score List [l] .v _{max of the} element in the List array, the processing routine is terminated.

In step S106, when the data acquisition unit 72 receives the latest data y _{Ln / Ty} via the data input unit 60, the pattern detection unit 76 performs steps similar to those in steps S102 to S105. The processing from S107 to step S110 is performed.

  Step S110 is realized by the processing routine shown in FIG. 8, and the processing of Step S130 to Step S136 is performed in the same manner as the processing of Step S120 to Step S126.

  In step S111, the pattern detection unit 76 determines whether or not each of the partial sequence pairs stored in the List array satisfies the condition shown in the above expression (10), and the condition shown in the above expression (10). If there is a partial sequence pair that satisfies the condition, the element is determined not to be replaced by a partial sequence pair that appears thereafter, and the corresponding score, start point, and end point are determined via the data output unit 80. The user is notified and the partial sequence pair is deleted from the List array.

  In step S112, the pattern detection unit 76 determines whether or not the reception of the two time-series data has been completed. If not completed (No), the process returns to step S101 or step S106, and if completed. (Yes), the process ends.

  As described above, according to the first embodiment of the present invention, all data of time-series data is used by calculating a similarity score using the sampling period of each time-series data. Even if there is little data, a similar partial sequence can be detected with high accuracy.

  In addition, in detecting similar patterns from data acquired by devices such as sensors, it is possible to detect similar patterns with accuracy close to that of original data while limiting the data transmitted by sensor nodes. It becomes. Further, the present invention can be applied not only to the sensor network but also to the case of detecting similar pattern detection from time series data at high speed.

[Second Embodiment]
<System configuration>
Next, a second embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The second embodiment is different from the first embodiment in that the sampling cycle is calculated on the sensor node side and the sampling cycle is periodically calculated.

  As shown in FIG. 9, the sensor node 200 according to the second embodiment includes a data sensing unit 10, a data acquisition unit 12, a sampling unit 14, a cycle storage unit 16, a wireless communication unit 18, and a cycle calculation unit 220. Composed.

  The period calculation unit 220 uses the information obtained from the data acquisition unit 12 to calculate the sampling period in an adversarial manner. The sampling period calculation method is the same as that of the period calculation unit 78.

  The sampling period calculated by the period calculation unit 220 is transmitted to the similar pattern detection device 150 by the wireless communication unit 18.

  The principle of detecting a similar partial sequence in the second embodiment will be described.

  In the second embodiment, the sampling cycle is updated as needed to detect similar patterns. In the first embodiment, a fixed sampling period is used. However, since the sampling period is determined using already acquired data, is it effective for data acquired thereafter? There is no guarantee. For example, consider data acquired from a temperature sensor. The data sequence that determines the sampling period is data that does not change much in temperature.If the temperature changes periodically due to changes in the weather or the environment after that, sampling is performed using the sampling period before the temperature change. A data sequence in which information on various changes is dropped is transmitted to the similar pattern detection apparatus 150 serving as a server. In order to cope with such a change, it is desirable to determine the sampling period and sample the data each time.

Therefore, in the present embodiment, the sampling period is appropriately recalculated in the sensor node 200 to appropriately adjust the sampling period. The original sequence acquired by the sensor node 200 is X = (x ₁ , ..., x _i , ..., x _n ), and the sequence sampled at the adjusted sampling period is X ′ = (x ₁ ,. .., x _i , ..., x _{n ′} ), and the sampling period is T _x = (t _x1 , ..., t _xi , ..., t _{xn ′} ). If the number of data between the latest data x _n currently acquired and the last data x _n′−1 transmitted from the sensor node 200 to the similar pattern detection device 150 is t _xn′−1 , the sensor node 200 The sensor node 200 transmits _xn to _{xn ′} to the similar pattern detection device 150 by the wireless communication unit 18. At the same time, the sampling period t _{xn ′} for the data to be transmitted next is determined using the original data up to x _n . The determined sampling period is sent to the similar pattern detection apparatus 150 for detecting a similar pattern.

  At this time, since it is impossible to hold all data acquired in the sensor node 200, it is desirable to discard past data and hold only the latest data for the calculation of the sampling period. For example, it may be possible to always maintain a certain number of data using a sliding window. In addition, an algorithm has been proposed that calculates the sampling period incrementally for the acquired data, instead of calculating the sampling period batchwise for the accumulated data (http://dl.acm.org/ citation.cfm? id = 1146470). By using these algorithms, the sampling period can be calculated without accumulating data in the sensor node 200.

The similar pattern detection apparatus 150 receives data from each sensor node 200 and information on the sampling period, and detects a similar pattern. Sampled sequence X ′ = (x ₁ , .., x _i , ..., x _{n ′} ) with latest element _{n ′} and sampled sequence Y ′ = with latest element m ′ Consider (y ₁ , ..., y _j , ..., y _{m ′} ). Also, the sampling period of the original sequence X of X is T _x = (t _x1 , ..., t _xi , ..., t _{xn ′} ), and the sampling period of the original sequence Y of Y is T _y = (t _y1 , ..., t _yj , ..., t _{ym ′} ). The score of the partial sequence pair in the sampled sequence is calculated according to the following equation (11).

Since the sequence length changes as the sampling period changes each time the score is calculated, the weights (B _v , B _h , B _d ) in each direction are also updated as needed. On the other hand, the position matrix is calculated according to the following equation (12) by incorporating the change of the sampling period into the starting point information.

<Operation of sensor node>
Next, the operation of the sensor node 200 according to the second embodiment will be described.

  A processing routine shown in FIG. 10 is executed by the sensor node 200.

First, in step S <b _> 201, the current data n from the data sensing unit 10 is acquired from the data acquisition unit 12. In step S202, the sampling unit 14 determines whether or not the interval between the previous sample element n′−1 and the current n has reached the previously calculated sampling cycle T _xn′−1 . When the interval between the element n′−1 of the previous sample and the current n has reached the previously calculated sampling period T _xn′−1 , in step S203, the period calculation unit 220 sets the interval up to x _n . The sampling period is calculated using the data and stored in the period storage unit 16. In step S204, the wireless communication unit 18 converts the data x _n acquired in step S201 (that is, the sampled data x _{n ′} ) and the sampling period calculated in step S203 to the similar pattern detection device. 150.

On the other hand, when the interval between the element n′−1 of the previous sample and the current time n has not reached the previously calculated sampling period T _xn′−1 , the process proceeds to step S205.

  In step S205, it is determined whether or not the reception of data from the data sensing unit 10 is completed. If not completed (No), the process returns to step S201, and if completed (Yes), the process is terminated.

<Operation of Similar Pattern Detection Device>
The similar pattern detection device 150 executes the similar partial sequence pair detection processing routine described in the first embodiment. Here, in step S103, the pattern detection unit 76 uses the similarity score v (n / T _x , j) of x _{n / Tx} and y _j and the start point s (n / T _x , j) to the above ( Calculation is performed using equations (11) and (12). Further, as the sampling period T _x, using a sampling period received from the sensor node 200 by the data input unit 60.

  In the same manner, step S108 is calculated using the same equations as the above equations (11) and (12).

  As described above, according to the second embodiment, all the data of the time series data is used by calculating the similarity score using the sampling period calculated as needed for each time series data. Therefore, even with a small amount of data, a similar partial sequence can be detected with high accuracy.

<Example>
In order to verify the effectiveness of the present invention, experiments using actual data were performed. In the experiment, a Linux (registered trademark) machine equipped with 4 GB of memory and an Intel (R) Core 2 2.4 GHz CPU was used. In order to visually grasp the experimental results, a scatter diagram is used here. The scatter diagram depicts the warping path of X [i _s : i _e ] and Y [j _s : j _e ], which are similar partial sequence pairs in the data sequences X and Y. That is, the element correspondence indicating which elements of the similar partial sequence pair are detected correspondingly is shown. In the scatter diagram, the horizontal axis represents the X element, and the vertical axis represents the Y element.

  FIG. 11 shows data acquired from the temperature sensor. A similar pattern that changes greatly between 18 degrees and 32 degrees is confirmed for data acquired at approximately one minute intervals. FIG. 12 shows the detection results when Examples 1 and 2 are used for the original similar pattern detection using Patent Document 3. As can be seen from the figure, both examples show results very close to the original detection.

  FIG. 13 is data in which the sunspot number is recorded. It is well known that sunspots have periodicity and are closely related to solar activity. When sun activity is active, many sunspots appear. Conversely, when sun activity is inactive, sunspots decrease. This change increases and decreases with a period of about 11 years. FIG. 14 shows the result of the detected pattern. This data also confirms the same result as the original similar pattern detection.

  15 and 16 show in detail the similar pattern detection result in the data set acquired from the temperature sensor, and FIG. 16 shows in detail the similar pattern detection result in the set in which the sunspot number is recorded. is there. Since the present invention is an approximation method and complements information on data that has not been transmitted to the server by sampling with information on the transmitted data, the DTW distance of the detected similar pattern is slightly different due to the difference in value. The result was different. However, as confirmed in FIGS. 12 and 14, it is confirmed that the position of the similar pattern can be detected with high accuracy.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

For example, although the square value of the Euclidean distance is used as ‖x _i −y _j 、, other distances such as the Euclidean distance may be used.

  In addition, specific configurations of hardware and software can be appropriately changed without departing from the gist of the present invention.

  Time series data is generated in various fields such as video, sensor networks, and finance. The present invention is applicable to all these fields.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Data sensing part 12, 72 Data acquisition part 14 Sampling part 16, 82 Period storage part 18 Wireless communication part 60 Data input part 70 Processing part 74 Detection condition storage part 76 Pattern detection part 78, 220 Period calculation part 80 Data output part 84 Matrix information storage unit 100, 200 sensor node 150 Similar pattern detection device

Claims (8)

From the first time series data obtained by sampling the first time series data at the first sampling period and the second time series data obtained by sampling the second time series data at the second sampling period, two similar partial sequence pairs are obtained. Similarity detected using a single score matrix whose matrix elements are DTW (Dynamic Time Warping) distances between subsequences and a similarity score that can be interconverted, and a position matrix whose matrix elements are the start positions of the partial sequences A partial sequence detection device comprising:
A storage unit for storing the score matrix and the position matrix;
When one element of any one of the sampled first time-series data and second time-series data is received, a portion in any one of the first time-series data and the second time-series data including the element Calculating a similarity score between the sequence and the partial sequence in the other one of the first time-series data and the second time-series data;
The calculated similarity score is stored as a matrix element of the score matrix of the storage unit corresponding to the end positions of the two partial sequences used for the calculation of the similarity score, and the similarity score is calculated. Storing the start positions of the two partial sequences used as matrix elements of the position matrix of the storage unit corresponding to the end positions;
A pair of partial sequences in which the similarity score stored in the score matrix of the storage unit is equal to or greater than a predetermined threshold is determined as a pair of similar partial sequences, and the pair of similar partial sequences is matched And a processing unit for detecting as
The processor is
When calculating any of the similarity scores,
The similarity score that is adjacent to the similarity score in the score matrix and that corresponds to the previous time with respect to the first time-series data is acquired, and the two portions of the object are obtained with respect to the acquired similarity score Subtract the value obtained by multiplying the first sampling period by the magnitude of the difference between the corresponding data elements in the sequence, and add the value obtained by multiplying a predetermined coefficient by a numerical value corresponding to the first sampling period. , Calculate the first score,
The similarity score that is adjacent to the similarity score in the score matrix and that corresponds to the previous time with respect to the second time-series data is acquired, and the two portions of the object are obtained with respect to the acquired similarity score Subtract the value obtained by multiplying the second sampling period by the magnitude of the difference between the corresponding data elements in the sequence, and add the value obtained by multiplying the predetermined coefficient by the numerical value corresponding to the second sampling period. To calculate the second score,
The similarity score that is adjacent to the similarity score in the score matrix and that corresponds to the previous time for both the first time-series data and the second time-series data is acquired, and the acquired similarity score On the other hand, a value obtained by multiplying the magnitude of the difference between corresponding data elements in the two target partial sequences by the larger one of the first sampling period and the second sampling frequency is subtracted, and the predetermined value is set. Adding a value obtained by multiplying the obtained coefficient by a numerical value corresponding to the sum of the first sampling period and the second sampling period to calculate a third score;
The similar partial sequence detection apparatus which uses the largest score among the first score, the second score, and the third score as the similarity score. The predetermined threshold is a value obtained by multiplying a predetermined value as a lower limit value of the length of two partial sequences of the pair of matching partial sequences by the predetermined coefficient;
The processor is
The similar partial sequence detection device according to claim 1, wherein when detecting a pair of matching partial sequences, the pair of matching partial sequences having a length equal to or longer than the lower limit value is detected by using the predetermined threshold. The processor is
Storing the detected pair of matching partial sequences in the storage unit;
When the detected matching partial sequence pair is stored in the storage unit, one of the matching partial sequence pairs stored in the storage unit has one matrix element in the position matrix used for calculation. If there is a match, a pair of matching partial sequences having the maximum similarity score among a plurality of matching partial sequence pairs having matching matrix elements in the position matrix is stored in the storage unit. The similar partial sequence detection device according to claim 2 for storing. The processor is
When the start positions of the two partial sequences used to calculate the similarity score are stored in the position matrix of the storage unit, the start positions of the two partial sequences used to calculate the acquired maximum similarity score Are stored in the position matrix of the storage unit as the start positions of the two partial sequences used for the calculation of the similarity score,
4. The method according to claim 1, wherein when the pair of matching partial sequences is detected, a start position corresponding to the similarity score equal to or higher than the predetermined threshold is specified with reference to a position matrix of the storage unit. The similar partial sequence detection device according to claim 1. From the first time series data obtained by sampling the first time series data at the first sampling period and the second time series data obtained by sampling the second time series data at the second sampling period, two similar partial sequence pairs are obtained. Detect using a single score matrix whose matrix element is a DTW (Dynamic Time Warping) distance between partial sequences and a similarity score that can be interconverted, and a position matrix whose matrix element is the start position of the partial sequence, A similar partial sequence detection method in a similar partial sequence detection device comprising a storage unit for storing the score matrix and the position matrix,
Depending on the processing unit,
When one element of any one of the sampled first time-series data and second time-series data is received, a portion in any one of the first time-series data and the second time-series data including the element Calculating a similarity score between the sequence and the partial sequence in the other one of the first time-series data and the second time-series data;
The calculated similarity score is stored as a matrix element of the score matrix of the storage unit corresponding to the end positions of the two partial sequences used for the calculation of the similarity score, and the similarity score is calculated. Storing the start positions of the two partial sequences used as matrix elements of the position matrix of the storage unit corresponding to the end positions;
A pair of partial sequences in which the similarity score stored in the score matrix of the storage unit is equal to or greater than a predetermined threshold is determined as a pair of similar partial sequences, and the pair of similar partial sequences is matched Including detecting as
When calculating any of the similarity scores by the processing unit,
The similarity score that is adjacent to the similarity score in the score matrix and that corresponds to the previous time with respect to the first time-series data is acquired, and the two portions of the object are obtained with respect to the acquired similarity score Subtract the value obtained by multiplying the first sampling period by the magnitude of the difference between the corresponding data elements in the sequence, and add the value obtained by multiplying a predetermined coefficient by a numerical value corresponding to the first sampling period. , Calculate the first score,
The similarity score that is adjacent to the similarity score in the score matrix and that corresponds to the previous time with respect to the second time-series data is acquired, and the two portions of the object are obtained with respect to the acquired similarity score Subtract the value obtained by multiplying the second sampling period by the magnitude of the difference between the corresponding data elements in the sequence, and add the value obtained by multiplying the predetermined coefficient by the numerical value corresponding to the second sampling period. To calculate the second score,
The similarity score that is adjacent to the similarity score in the score matrix and that corresponds to the previous time for both the first time-series data and the second time-series data is acquired, and the acquired similarity score On the other hand, a value obtained by multiplying the magnitude of the difference between corresponding data elements in the two target partial sequences by the larger one of the first sampling period and the second sampling frequency is subtracted, and the predetermined value is set. Adding a value obtained by multiplying the obtained coefficient by a numerical value corresponding to the sum of the first sampling period and the second sampling period to calculate a third score;
The similar partial sequence detection method which uses the maximum score among the first score, the second score, and the third score as the similarity score. The predetermined threshold is a value obtained by multiplying a predetermined value as a lower limit value of the length of two partial sequences of the pair of matching partial sequences by the predetermined coefficient;
By the processing unit,
6. The similar partial sequence detection method according to claim 5, wherein when detecting the matching partial sequence pair, the matching partial sequence pair having a length equal to or longer than the lower limit value is detected by using the predetermined threshold. By the processing unit,
Storing the detected pair of matching partial sequences in the storage unit;
When the detected matching partial sequence pair is stored in the storage unit, one of the matching partial sequence pairs stored in the storage unit has one matrix element in the position matrix used for calculation. If there is a match, a pair of matching partial sequences having the maximum similarity score among a plurality of matching partial sequence pairs having matching matrix elements in the position matrix is stored in the storage unit. The similar partial sequence detecting method according to claim 6, wherein the similar partial sequence detecting method is stored.   The program for functioning a computer as each part of the similar partial sequence detection apparatus of any one of Claims 1-4.