KR20100097951A - Method and apparatus for classifying multivariate stream data - Google Patents

Abstract

PURPOSE: A method and an apparatus for classifying multivariate stream data are provided to accurately classify multivariable stream data by extracting and using the important features from stream data inputted from sensors. CONSTITUTION: A data converter(310) uses a symbol to convert inputted multivariable stream data into one character string. A partial character string generating unit(320) applies an n-gram scheme to the converted character string to create the set of partial character strings having an n number of syllables. A motif extractor(330) extracts a partial character string from the set of the partial character string, wherein the partial character string can be the motif for each class. A data classifier(360) classifies the multi-variable stream data into one among class sets based on the partial character strings.

Description

METHOD AND APPARATUS FOR CLASSIFYING MULTIVARIATE STREAM DATA}

The present invention relates to a method and apparatus for classifying multivariate stream data taking into account the temporal relationship between motifs.

Stream data classification is to classify new stream data into one of a predefined class set. Most existing classification techniques choose the most similar class by considering the numerical distance and statistical distribution of the data. However, since multivariate stream data has a pattern having strong associations and sequential characteristics between attributes, the classification technique using the similarity between simple distance comparison and statistical distribution is limited.

The stream data pattern classification technique is very useful for finding patterns in historical data collected by sensors and classifying new data using these patterns. Unlike traditional time series data such as stocks, weather, and population data, advances in sensors and wireless network technologies have enabled real-time data collection in real time. This data collection technology has expanded the need for users to not only monitor sensor data values, but also to characterize current data and predict future conditions.

1 is a diagram illustrating an example of stream data.

For example, a mobile robot equipped with several sensors may perform a task and transmit measured values collected from each sensor to a central server at regular intervals. As robots perform their jobs, they may face situations such as rotation, pinch, collision, and obstacles. In addition, depending on the situation, each sensor can collect values that do not increase, decrease or change rapidly over time. If the user can see the multivariate stream data obtained from the plurality of sensors and can accurately classify the performance patterns of the distant robots, the robots can be controlled or predict the state of the robots in the future.

On the other hand, existing stream data classification methods include a distance-based classification method for selecting the nearest object using a distance measure, a classification method for selecting the most similar object using statistical information, and a classification method using structural information.

The distance-based classification technique selects the closest distance object using the distance measure of the numerical vector of each attribute column. The most common distance scale technique is the Euclidean distance scale, or Dynamic Time Wrapping (DTW), which selects the class with the closest total sum.

The classification method using statistical information is a method using probabilistic theory and data distribution characteristics such as Bayesian classifier, HMM, RNN. These techniques use the previously learned probability values and distribution characteristics to select the maximum likelihood class.

Finally, the structure pattern classification technique is to create a rule, make the characteristics of the data into a tree or graph, and select a class that follows the most similar structure.

However, in the sensor network application, stream data collected through various sensors has unique data characteristics for each class, and because there is a strong temporal causal relationship between the attributes of the stream data, classification methods using only simple distance, probability and structure are multivariate. Not suitable for classifying stream data.

An embodiment of the present invention extracts important features that are differentiated from other classes of stream data from stream data input from a plurality of sensors, and considers the importance of the extracted features in a class and the time relationship pattern between the features. To provide a more accurate classification of multivariate stream data.

As a technical means for achieving the above-described technical problem, a first aspect of the present invention is a data conversion unit for converting the input multi-variable stream data into a single character string using a symbol, an n-gram (n- a substring generator for generating a set of substrings having n syllables by applying a gram method, and a part capable of becoming a motif for each class in the generated substring set A multivariate stream data including a motif extractor for extracting a character string and a data classifier for classifying the multivariate stream data into one of the class sets based on the substring that may be the extracted motif as one of a predefined class set. A device for sorting can be provided.

In addition, the second aspect of the present invention is to (a) converting the input multivariate stream data into a single character string using a symbol, (b) applying the n-gram method to the converted character string Generating a set of substrings having n syllables (n is a natural number), and (c) extracting a substring that can be a motif for each class from the generated set of substrings And classifying the multivariate stream data into one of the class sets based on the substring that may be the extracted motif. Can provide.

According to the above-described problem solving means of the present invention, by simplifying the numerical-based multivariate stream data using symbols associated with the degree of change of the data, the complexity of the interpretation algorithm of the stream data is reduced, Because of the use of symbols, an approximate interpretation of the generated rule is possible.

In addition, according to the above-described problem solving means of the present invention, by using the n-gram method, the problem of the conventional stream data classification method that can compare the similarity only if the length of the training data and the test data can be solved. have.

In addition, according to one of the other problem solving means of the present invention, since the motif is selected and applied to the classification algorithm, space and time cost can be saved than considering the entire data.

In addition, according to one of the other problem solving means of the present invention, since the motif was selected based on the TFIDF value considering the frequency of occurrence in another class, rather than simply selecting the motif with high frequency in one class, Stream data can be classified.

In addition, according to another means for solving the problem of the present invention, the rule analysis is possible because the probability value of the motif, the TFIDF value, and the mutual information value are considered together and the probability value of the data and the structural pattern are considered. For example, stream data can be classified more accurately.

In addition, according to another means for solving the problem of the present invention, it is possible to analyze behavior using robot data, recognition of sign language, event analysis and classification using bio biodata, and analyze multivariate stream data collected in RFID and USN fields. have.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element in between. . In addition, when a part is said to "include" a certain component, which means that it may further include other components, except to exclude other components unless otherwise stated.

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

2 is a schematic diagram of a multivariate stream data classification system according to an embodiment of the present invention.

In the present specification, in addition to the x, y, and z directions, each person wears gloves with both sensors attached to each of the eleven sensors that can detect motions such as roll, pitch, and yaw. The multivariate stream data obtained when the word 'is expressed in sign language will be described as an example.

When a person wears gloves with 11 sensors on both hands and expresses the word 'YES' in sign language, each sensor may collect data at regular time intervals and transmit the data to the multivariate stream data classification apparatus 300. .

The multivariate stream data transmitted to the multivariate stream data classification apparatus 300 may be composed of values that do not increase, decrease or change rapidly over time while expressing the word 'YES'. In addition, since multivariate stream data is collected from 22 sensors, the number of streams is 22. That is, 22 multivariate stream data can be obtained while representing a class called 'YES' in sign language.

The apparatus for classifying multivariate stream data 300 may convert the multivariate stream data based on the numerical value of each class into a single character string using a symbol. In this case, the symbol is a symbol associated with the degree of change in the data between two time points of the multivariate stream data.

For example, the multivariate stream data classification apparatus 300 may convert the multivariate stream data into one string using five symbols. The symbol may be U, u, D, d, S. Where U is a sudden rise in data between two points, u is a monotonous rise in data between two points, D is a sharp decrease in data between two points, d is a monotonous decrease in data between two points, and S is stabilization. it means.

In addition, the multivariate stream data classification apparatus 300 may generate a partial string set by applying an n-gram method to the converted string. Here, the en-gram method divides one long string into n adjacent syllable units while maintaining the order. For example, if you apply the 2-gram method to 'South Korea', the substrings of 'Korea', 'Hanmin' and 'South Korea' are generated.

Also, the multivariate stream data classification apparatus 300 may extract a substring that may be a motif for each class from the generated substring set. Motif here refers to an event that can be an important feature of each class. For example, the multivariate stream data classification apparatus 300 may extract a character string that appears only when the word 'YES' is expressed among substrings of the multivariate stream data obtained when the word 'YES' is expressed in sign language. Can be.

In addition, the multivariate stream data classification apparatus 300 generates time relationship information between any two substrings within a preset time range among substrings that may be the extracted motif, and generates the time relationship information. A weight value occupied by each class may be calculated.

As described above, according to the multivariate stream data classification apparatus 300, not only the recognition of sign language using stream data, but also behavior analysis using robot data, event analysis and classification using bio biometric data, radio frequency identification (RFID), USN, etc. In all fields using stream data collected by multivariate such as (Ubiquitous Sensor Network), a specific event may be detected, or states and actions may be classified using the detected event.

The detailed configuration of the multivariate stream data classification apparatus 300 and the function of each configuration will be described in detail with reference to FIG. 3.

3 is a detailed block diagram of a multivariate stream data classification apparatus according to an embodiment of the present invention.

As shown, the multivariate stream data classification apparatus according to an embodiment of the present invention includes a data converter 310, a substring generator 320, a motif extractor 330, a time information generator 340, and a mutual And an information generator 350 and a data classifier 360.

The data converter 310 may convert the numerical-based multivariate stream data for each class into a single string using a symbol associated with the degree of change in the data between two viewpoints of the multivariate stream data.

The data converter 310 may normalize the numerical based multivariate stream data in order to convert the multivariate stream data into a character string using a symbol. Data normalization is the conversion of values between -1 and 1 in different ranges for different properties. Normalization allows you to compare values between attributes and see how far each successive attribute value is from the mean.

In addition, the data transformer 310 calculates a cumulative probability distribution for the difference between normalized values of two consecutive time points of the numerical-based multivariate stream data, and breakpoints based on the calculated cumulative probability distribution. You can decide. The break point may be a boundary point for determining the type of symbol.

For example, the symbol may be U, u, D, d, s. Where U is a sudden rise in data between two points, u is a monotonous rise in data between two points, D is a sharp decrease in data between two points, d is a monotonous decrease in data between two points, and S is stabilization. it means. In other words, if the difference between the normalized values of two consecutive time points is -0.25 ~ +0.25, S, if the difference between the normalized values of two consecutive time points is +0.25 ~ +0.84, u, the normalization of two consecutive time points If the difference between the normalized values is +0.84 or more, U, the difference between the normalized values of two successive time points is -0.25 to -0.84, and the difference between the normalized values of the two successive time points is -0.84 or less. Can be used.

4 illustrates multivariate stream data converted into character strings according to an embodiment of the present invention.

As shown, one stream after normalization can be converted to a string using symbols. That is, the data converter 310 may convert one stream data into one long string using five symbols.

For example, the converted string is' 16DsDU,... , sss ”. Here, the number 16 means the attribute of the 16th sensor among the 22 sensors. In other words, this property refers to the measurements that are collected when the palm is shaken in one direction. Also, ‘UsDU,… , sss' means that as time goes by, the data value changes from “Rapid Rise (U) → Stabilization (s) → Rapid Decrease (D) → Rapid Rise (U),…. , → stabilization (s) ”.

3, the substring generator 320 applies the n-gram method to the string generated by the data converter 310, and has a substring having n syllables (n is a natural number). You can create a set. For example, the substring generator 320 may include a set of substrings having two syllables, a set of substrings having three syllables, a set of substrings having four syllables, and a substring having five syllables. You can create one or more sets of sets.

Because stream data has locality between successive values, a set of multiple substrings can represent data characteristics better than one long string. In addition, substrings can be compared in long strings by splitting them into one sequenced long string.

5 is a diagram illustrating a partial string set according to an embodiment of the present invention.

For example, as illustrated, a partial string set may be generated by applying the 1-gram method or the 5-gram method to 22 strings converted using 5 symbols. In other words, the substring generator 320 generates a substring set by applying the 1-gram method or the 5-gram method to 22 strings received from the data converter 310.

3, the motif extractor 330 may extract a substring that may be a motif for each class from the substring set generated by the substring generator 320. In order to improve the quality and speed of data classification, it is essential to use an important feature representing the class, motifs. There are two problems with classification techniques that process huge data. First, it requires a lot of execution time because the data to be processed is too large. Second, it is difficult to strictly determine the characteristics that represent each class. When analyzing large documents with hundreds of features, or sensor data collected simultaneously from dozens of sensors, selecting and analyzing important features will greatly improve the accuracy and speed of data classification.

The motif extractor 330 may use a TFIDF (Term Frequency and Inverse Document Frequency) value to extract a substring that may be a motif. TFIDF is a technique for determining the weight of each substring for each class. The TFIDF value is determined by the following equation.

In Equation 1

는 발생 빈도로서 클래스 또는 문서(d) 안에서 단어(t)가 발생한 횟수를 의미한다.

Is the frequency of occurrence and means the number of occurrences of the word t in the class or document d.

는 역문서 빈도로서 다음 수학식에 의해 결정된다.Also,

Is the inverse document frequency, which is determined by the following equation.

In equation (2)

는 문서 빈도로서, 특정한 단어(t)가 적어도 한번 발생한 문서의 개수를 의미한다. 또한,

는 문서의 총 개수를 의미한다. 그러므로 많은 문서에 특정한 단어(t)가 포함되어 있다면 역문서 빈도 값은 낮고, 특정한 문서에만 상기 특정한 단어(t)가 포함되어 있다면 역문서 빈도 값은 높다.

Is the document frequency, and means the number of documents in which a specific word t occurs at least once. Also,

Means the total number of documents. Therefore, if many documents contain a specific word t, the reverse document frequency value is low, and if only a specific document contains the specific word t, the reverse document frequency value is high.

Since the TFIDF value is calculated as the product of the occurrence frequency value and the reverse document frequency value, the specific substring appears more in a specific class, but the lower the substring appears in another class, the higher the TFIDF value appears. Therefore, a substring that appears a lot in a specific class but appears less in another class may be a motif of the specific class.

6 is a TFIDF value table for a substring according to an embodiment of the present invention.

As shown, '22uud' has a high weight value only in the 'boy' class and the 'come' class, and thus may be a motif representing the 'boy' class and the 'come' class. That is, if sensor 22 has a monotonous rise (u) → monotonous rise (u) → monotonous decrease (d), it can be predicted that the word 'boy' or 'come' is displayed.

However, if the value of the weight is high or zero in all classes such as '5dsud', it cannot be a motif representing the class.

In addition, when the TFIDF value is obtained for all the substrings generated by the substring generator 320, the TFIDF value for the substring using the 1-gram method is almost zero, and the substring to which the 1-gram method is applied is a motif. It cannot be a value. On the other hand, substrings applying the 2 to 5-gram method are most likely to be motifs.

3, the time information generator 340 may generate time relationship information between any two substrings among substrings that may be motifs extracted by the motif extractor 330.

7 is a time relationship information table according to an embodiment of the present invention.

As shown, the motif is time interval data having a start point and an end point.

In addition, the time relationship information may be Finish, Processing, Start, Overlap, Meet, and Before. Finish is time relationship information in which end points of the first motif and the second motif coincide with each other. During is temporal relationship information in which the first motif is included in the second motif. At this time, the second motif may be included in the first motif. Start is time relationship information in which the start time of the first motif and the second motif coincide. Overlap is time relationship information in which the first motif and the second motif overlap for some time. Meet is time relationship information where the second motif starts immediately after the first motif ends. At this time, it does not matter even if the first motif starts immediately after the second motif ends. Finally, Before is the time relationship information where the first motif ends and the second motif starts, but there is no temporal continuity.

The time information generator 340 may display each time relationship information as six characters B, M, O, S, D, and F. In addition, for example, the time information generator 340 may display the time relationship between the motifs in the form of {first motif (time relationship information character) second motif}.

The mutual information generator 350 may calculate a weight value of each class of time relationship information within a preset time range among time relationship information generated by the time information generator 340. .

Even though the time information generator 340 generates time relationship information between all the motifs for the entire time, since there is a strong association between adjacent neighboring motifs, the mutual information generator 350 may have a constant time. Only time relationship information between the motifs in the range d may be considered.

8 is a diagram illustrating time relationship information between motifs according to an embodiment of the present invention.

For example, the time information generator 340 may detect that the motif '5uuud' includes the motif '4ud' and generate time relationship information called {5uuud (D) 4ud}. In addition, the time information generator 340 may discover that the motif '5uuud' and the motif '3Ddsd' overlap and generate time relationship information called {5uuud (O) 3Ddsd}. In addition, the time information generator 340 may discover that the motif '4ud' and the motif '3Ddsd' overlap and generate time relationship information called {4ud (O) 3Ddsd}. Also, the time information generator 340 may generate time relationship information such as {3Ddsd (M) 2Ddsd}, {4Ddsd (S) 3ds}, {2Ddsd (D) 4Ddsd}, and {2Ddsd (D) 3ds}. .

However, the mutual information generator 350 may consider only time relationship information between motifs within a predetermined time range d.

The mutual information generator 350 may use a mutual information value to calculate a weight value of each class of time relationship information between two motifs. In other words, the mutual information generating unit 350 mutually intersects any two substrings within a preset time range among the substrings extracted by the motif extracting unit 330 and the time relationship information formed by the two substrings. A weight value that occupies the time relationship information formed by the two substrings in each class may be calculated using an information value.

The mutual information value is determined by the following equation.

과

는 각각 모티프(

)과 모티프(

)가 특정 클래스의 모티프들 중에서 차지하는 비율이다. 또한,

는 모티프(

)과 모티프(

)가 미리 설정된 시간 범위(d) 내에서 형성하는 시간 관계 정보가 특정 클래스의 모티프들 중에서 차지하는 비율이다.here

and

Is a motif (

) And motifs (

) Is the percentage of motifs of a particular class. Also,

Is the motif (

) And motifs (

) Is a ratio of time-related information formed within a preset time range (d) among motifs of a specific class.

9 is a table of mutual information values for time relationship information according to an embodiment of the present invention.

As shown, the temporal relationship information M formed by the motifs '2uud' and '5uds' has a high mutual information value, that is, a weight value, for the class 'girl'.

In other words, for example, if sensor # 2 maintains a monotonous rise, monotonous decrease, and stabilization immediately after sensor # 2 monotonically increases → monotonically rises → monotonically decreases, it can be predicted that “girl” is expressed in sign language.

In addition, the time relationship information O formed by the motifs '2udu' and '5uds' has a high weighting value for the class 'hello'.

In other words, if sensor # 2 maintains monotonous rise, monotonous decrease, and stabilization while sensor # 2 is monotonically rising → monotonically decreasing → monotonically rising, it can be predicted that “hello” is expressed in sign language.

3, the data classifier 360 may classify the multivariate stream data into one of a predefined class set based on a substring that may be a motif extracted by the motif extractor 330. Further, the data classifier 360 may classify the multivariate stream data into one of a predefined class set in consideration of the time relationship information between the substrings generated by the time information generator 340. In addition, the data classifying unit 360 may classify the multivariate stream data into one of a predefined class set based on a weight value of each class of the time relationship information calculated by the mutual information generating unit 350.

As described above, the apparatus for classifying multivariate stream data according to an embodiment of the present invention discovers important motifs representing each class and considers the time relationship therebetween, thereby enabling more accurate data interpretation and prediction.

10 is a flowchart of a multivariate stream data classification method according to an embodiment of the present invention.

In operation S1000, the data converter 310 may convert the input multivariate stream data into a single character string using a symbol associated with a degree of change in the data between two viewpoints of the multivariate stream data. For example, the symbol may be U, u, D, d, s.

In operation S1000, the data converter 310 may normalize the numerical based multivariate stream data in order to convert the multivariate stream data into a character string using a symbol.

In operation S1000, the data converter 310 calculates a cumulative probability distribution for the difference between normalized values of two consecutive time points of the numerical based multivariate stream data, and based on the calculated cumulative probability distribution. You can determine the breakpoint. The break point may be a boundary point for determining the type of symbol.

In operation S1020, the substring generator 320 applies an n-gram method to the string converted in operation S1000 to generate a substring set having n syllables (n is a natural number). Generating step. For example, in step S1020, the substring generator 320 may include a set of substrings having two syllables, a set of substrings having three syllables, a set of substrings having four syllables, and five syllables. One or more sets of sets of substrings with

In operation S1040, the motif extracting unit 330 extracts a substring that may be a motif for each class from the substring set generated in operation S1020.

In operation S1040, the motif extractor 330 may use a TFIDF (Term Frequency and Inverse Document Frequency) value to extract a substring that may be a motif. In a specific class, a large number of specific substrings appear, but in other classes, the smaller the substring, the higher the TFIDF value. Therefore, a substring that appears a lot in a specific class but appears less in another class may be a motif of the specific class. Formal description of the TFIDF value is the same as described above will be omitted.

In step S1060, the time information generator 340 generates time relationship information between any two motifs among the motifs extracted in the step S1040.

For example, the time relationship information may be Before, Meet, Overlap, Start, During, and Finish, each time relationship information being 6 pieces. The letters B, M, O, S, D, and F may be represented.

In step S1080, the mutual information generating unit 350 calculates a weight value of each time relationship information within a predetermined time range among the time relationship information between the motifs generated in step S1060. It's a step.

Although the time information generator 340 generates time relationship information between all the motifs for the entire time in step S1060, the mutual information generator 350 may determine the time between the motifs within a certain time range d. Only relationship information can be considered.

In operation S1080, the mutual information generator 350 may use a mutual information value to calculate a weight value of time relation information between two motifs in each class. Since the formal description of the mutual information value is the same as described above, it will be omitted.

In operation S1100, the data classifying unit 360 classifies the input multivariate stream data into one of a predefined class set. For example, in operation S1100, the data classifier 360 may classify the multivariate stream data into one of a predefined class set based on the substring that may be the motif extracted in operation S1040. Furthermore, in operation S1100, the data classifier 360 may classify the multivariate stream data into one of a predefined class set in consideration of the time relationship information between the substrings generated in operation S1060. In addition, the data classifying unit 360 may classify the multivariate stream data into one of a predefined class set based on a weight value of each class of the time relationship information calculated in operation S1080.

The apparatus and method for classifying multivariate stream data according to an embodiment of the present invention collects multivariate variables such as behavior analysis using robot data, event analysis and classification using biometric data, RFID, USN, as well as recognition of sign language using stream data. It can be used to detect specific events in all fields using stream data, or to classify states and actions using the detected events.

In addition, the apparatus and method for classifying multivariate stream data according to an embodiment of the present invention considers only a motif, thereby saving spatial and time costs than considering the entire data. In addition, the temporal relationship between the motifs is reflected in the rules to enable clear interpretation of the classification rules.Because the probability and time structure information are considered together, the data can be classified accurately with existing techniques that simply consider data distance or data distribution. .

One embodiment of the present invention can also be implemented in the form of a recording medium containing instructions executable by a computer, such as a program module executed by the computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transmission mechanism, and includes any information delivery media.

The foregoing description of the present invention is intended for illustration, and it will be understood by those skilled in the art that the present invention may be easily modified in other specific forms without changing the technical spirit or essential features of the present invention. will be. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is shown by the following claims rather than the above description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention. do.

1 is a diagram illustrating an example of stream data.

2 is a schematic diagram of a multivariate stream data classification system in accordance with an embodiment of the present invention.

3 is a detailed block diagram of an apparatus for classifying multivariate stream data according to an embodiment of the present invention.

4 illustrates multivariate stream data converted to character strings according to an embodiment of the present invention.

5 illustrates a partial string set according to an embodiment of the present invention.

6 is a table of TFIDF values for substrings in accordance with one embodiment of the present invention;

7 is a time relationship information table according to an embodiment of the present invention.

8 illustrates time relationship information between motifs according to an embodiment of the present invention.

9 is a table of mutual information values for time relationship information according to an embodiment of the present invention.

10 is a flowchart of a method for classifying multivariate stream data according to an embodiment of the present invention.

Claims (23)

An apparatus for classifying multivariate stream data into one of a predefined class set, A data converter for converting the input multivariate stream data into a single string using symbols; A substring generator for generating a set of substrings having n syllables (n is a natural number) by applying an n-gram method to the converted string; A motif extracting unit that extracts a substring that may be a motif for each class from the generated substring set; and A data classification unit classifying the multivariate stream data into one of the class sets based on the extracted substring Multivariate stream data classification device comprising a. The method of claim 1, A time information generator for generating time relationship information between any two substrings of the extracted substrings; Further comprising a mutual information generating unit for calculating a weight value occupied by each class of the time relationship information within a preset time range, And the data classifying unit classifies the multivariate stream data into one of the class sets based on the time relationship information and the weight value. The method of claim 1, The data converter normalizes the multivariate stream data, calculates a cumulative probability distribution for the difference between normalized values of two consecutive time points of the multivariate stream data, and breakpoints based on the calculated cumulative probability distribution. And determine the preference based on the breakpoint. The method of claim 1, The substring generator generates at least one set of substrings having two syllables, a set of substrings having three syllables, a set of substrings having four syllables, and a set of substrings having five syllables. Multivariate stream data classification device. The method of claim 3, wherein And wherein the preference is associated with a degree of change in the data between two time points of the multivariate stream data. The method of claim 3, wherein The symbol includes U, u, D, d and S, Wherein U is a sudden rise in data between two points in time of the multivariate stream data, u is a monotonous rise, D is a sharp decrease, d is a monotonous decrease, and S is a stabilization. The method of claim 3, wherein The symbol includes U, u, D, d and S, S is the difference between the normalized value of the two successive time points is -0.25 ~ +0.25, when u is the difference between the normalized value of the two successive time points is +0.25 ~ +0.84, the U is If the difference between the normalized value of two consecutive time points is +0.84 or more, the d is the difference between the normalized value of the two consecutive time points is -0.25 ~ -0.84, the D is normalized of the two consecutive time points A multivariate stream data classification apparatus that is used when the difference in values is -0.84 or less. The method of claim 1, And the motif extracting unit uses a term frequency and inverse document frequency (TFIDF) value that determines a weight of each substring for each class. The method of claim 2, And the temporal relationship information includes a before, a meet, an overlap, a start, a containing, and a finish. The method of claim 9, The mutual information generating unit using a mutual information value between any two substrings within a preset time range among the extracted substrings and the time relationship information formed by the two substrings Multivariate stream data classification device. In the method of classifying multivariate stream data into one of a predefined class set, (a) converting the input multivariate stream data into a single string using symbols; (b) generating a set of substrings having n syllables (n is a natural number) by applying an n-gram method to the converted string; (c) extracting a substring that can be a motif for each class from the generated substring set; and (d) classifying the multivariate stream data into one of the class sets based on the substring that may be the extracted motif. Multivariate stream data classification method comprising a. The method of claim 11, (e) generating time relationship information between any two of the extracted substrings; and (f) calculating a weight value occupied by the respective classes by the time relationship information within a preset time range; And (d) classifying the multivariate stream data into one of the class sets based on the time relationship information and the weight value. The method of claim 11, Step (a) is (a1) normalizing the multivariate stream data, (a2) calculating a cumulative probability distribution for the difference between normalized values of two consecutive time points of the multivariate stream data; and (a3) determining a breakpoint based on the calculated cumulative probability distribution The multivariate stream data classification method comprising a. The method of claim 11, In step (b), at least one of a set of substrings having two syllables, a set of substrings having three syllables, a set of substrings having four syllables, and a set of substrings having five syllables Steps to generate Multivariate stream data classification method comprising a. The method of claim 13, And wherein the preference is associated with a degree of change in the data between two time points of the multivariate stream data. The method of claim 13, The symbol includes U, u, D, d and S, Wherein U is a sudden rise in data between two time points of the multivariate stream data, u is a monotonous rise, D is a sharp decrease, d is a monotonous decrease, and S is a stabilization. The method of claim 13, The symbol includes U, u, D, d and S, Where S is the difference between the normalized values of the two successive time points is -0.25 to +0.25, and u is the difference between the normalized values of the two successive time points is +0.25 to +0.84, and U is When the difference between the normalized value of the two successive time points is +0.84 or more, d is the difference between the normalized value of the two successive time points is -0.25 ~ -0.84, D is the normalization of the two successive time points The method of classifying multivariate stream data which is used when the difference between the calculated values is -0.84 or less. The method of claim 11, The step (c) is to use a TFIDF (Term Frequency and Inverse Document Frequency) value for determining the weight of each substring for each class. 13. The method of claim 12, And the temporal relationship information includes before, meet, overlap, start, containing and finish. The method of claim 19, Step (f) uses a mutual information value between any two substrings within a preset time range among the extracted substrings and the time relationship information formed by the two substrings. Multivariate stream data classification method. A computer-readable recording medium having recorded thereon a program for performing the steps according to any one of claims 11 to 20. An application device for performing an application operation using multivariate stream data classified according to any one of claims 1 to 10. An application apparatus for performing an application operation using multivariate stream data classified according to any one of claims 11 to 20.

Images