KR20200052425A - Method for analyzing time series data, determining a key influence variable and apparatus supporting the same - Google Patents

Abstract

Provided is a method of determining a key influence variable for a target class among a plurality of time series variables. The method of determining a key influence variable comprises the following steps of: acquiring a first matrix predicted as the target class from multiple time series data related to the time series variables; acquiring a second matrix belonging to a specific class; calculating similarity between the two matrixes except a value of a first time series variable from the first matrix and the second matrix; and determining the first time series variable as the key influence variable in response to determination that the calculated similarity satisfies a predetermined condition, wherein a first row or a first column of the first matrix may be formed of a measured value of the first time series variable and a second row or a second column of the first matrix may be formed of a measured value of a second time series variable.

Description

Methods for analyzing time series data, methods for determining main influence variables, and devices that support the methods {METHOD FOR ANALYZING TIME SERIES DATA, DETERMINING A KEY INFLUENCE VARIABLE AND APPARATUS SUPPORTING THE SAME}

The present invention relates to a method for analyzing time series data, a method for determining a main influence variable, and an apparatus supporting the methods. More specifically, in order to improve the accuracy of data analysis, a method of analyzing multiple time series data in consideration of a correlation between a plurality of time series variables, and a method of determining a main influence variable having the most influence on an analysis result among a plurality of time series variables And devices that support the methods.

The semiconductor process is one of the most complex manufacturing processes and most of the processes are automated. An anomaly detection technology is one of the essential technologies for efficiently operating an automated semiconductor process. This is because if the failure occurs in an automated process, the entire manufacturing process is interrupted and the amount of economic loss increases rapidly.

In a semiconductor process, multiple sensors are used to detect an abnormal condition in real time, and a large amount of data is generated in real time from multiple sensors. This multi-time series data consists of measurements for as few as tens and as many as hundreds of time series variables (e.g. temperature, humidity, etc.).

The conventional anomaly detection method is a method of analyzing characteristics of time series data for each time series variable and predicting the presence or absence of an abnormality according to the analysis result. That is, although multiple time series variables are being monitored, analysis for abnormality detection was performed independently of each other based on a single time series variable. Accordingly, the correlation between time series variables was not reflected in the anomaly detection process, and as a result, there was a problem in that the accuracy of the anomaly detection was deteriorated.

Some methods have been proposed to perform anomaly detection considering the correlation of a small number of time series variables, but a method to consider the correlation between tens and hundreds of time series variables has not been proposed.

In addition, a method for accurately identifying the main influence factor that has the most influence on the abnormal state among the multiple time series variables has not been proposed.

Korean Patent Publication No. 10-2016-0026492 (2016.03.09 published)

The technical problem to be solved by the present invention is to provide a method for analyzing multiple time series data in consideration of a correlation between a plurality of time series variables and an apparatus supporting the method in order to improve analysis accuracy.

Another technical problem to be solved by the present invention is to provide a method for accurately determining a main influence variable having the most influence on an analysis result among the plurality of time series variables and an apparatus supporting the method.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

In order to solve the above technical problem, a method for determining a main influence variable according to an embodiment of the present invention includes: a method for determining a main influence variable for a target class among a plurality of time series variables in a computing device; In multiple time series data associated with a plurality of time series variables, obtaining a first matrix predicted by the target class, obtaining a second matrix belonging to a specific class, and a first time series in the first matrix and the second matrix Excluding the value of the variable, calculating the similarity between the two matrices and determining the first time-series variable as the main influence variable in response to determining that the calculated similarity satisfies a predetermined condition. have. In this case, the first row or the first column of the first matrix may consist of measured values of the first time series variable, and the second row or the second column of the first matrix may consist of measured values of the second time series variable. have.

In one embodiment, the specific class is a class different from the target class,

The determining of the main influence variable may include determining the first time series variable as the main influence variable in response to a determination that the calculated similarity is equal to or greater than a threshold value.

In one embodiment, the specific class is the same class as the target class, and the determining of the main influence variable comprises: in response to the determination that the calculated similarity is less than a threshold value, the first time series variable is the main influence. It may include the step of determining as a variable.

In one embodiment, the step of acquiring the first matrix comprises: extracting data of a predetermined time series section from the multi-time series data to generate the first matrix, and based on an analysis result of the first matrix. It may include the step of predicting the class of one matrix to the target class.

In one embodiment, the obtaining of the second matrix may include obtaining a plurality of candidate matrices belonging to the specific class, and selecting a second matrix from the plurality of candidate matrices by applying a locality sensitive hashing (LSH) algorithm. It may include the steps.

A method for determining a main influence variable according to another embodiment of the present invention for solving the above technical problem is a method for determining a main influence variable for a specific class among a plurality of time series variables in a computing device, wherein the plurality of time series variables Generating a first matrix by extracting data of a preset time series section from multiple time series data associated with, inputting the first matrix to a prediction model, and based on a first confidence score output from the prediction model And predicting a class of the first matrix as a first class and determining a main influence variable for the first class, wherein a first row or a first column of the first matrix is a first time series variable. The second row or the second column of the first matrix is composed of the measured values of the second time series variable, The determining of the main influence variable may include re-entering the first matrix from which the value of the first time series variable is excluded, into the prediction model to obtain a second confidence score, and wherein the second confidence score is predetermined. And in response to the determination that the condition is satisfied, determining the first time series variable as the main influence variable of the first class.

In one embodiment, the first confidence score is a confidence score for the first class, the second confidence score is a confidence score for the second class, and the first time series variable is a major influence variable of the first class The determining may include determining the first time-series variable as a primary influence variable of the first class in response to determining that the second confidence score is equal to or greater than a threshold value.

In one embodiment, the first confidence score and the second confidence score are both confidence scores for the first class, and determining the first time series variable as the primary influence variable of the first class includes: And determining the first time series variable as the main influence variable in response to a determination that a difference between the first confidence score and the second confidence score satisfies a predetermined condition.

The time series data analysis method according to another embodiment of the present invention for solving the above technical problem analyzes multiple time series data associated with a prediction target using a convolutional neural network-based prediction model in a computing device In the method, extracting data of a preset time series section from the multiple time series data, generating a first matrix, and applying the first matrix to the prediction model to predict a class of the prediction target It can contain. In this case, the multiple time series data includes measured values of the first time series variable and the second time series variable, and the first row or the first column of the first matrix is the measured value of the first time series variable for the time series section. The second row or the second column of the first matrix may be configured as measured values of the second time series variable for the time series section.

In one embodiment, the step of generating the first matrix comprises measuring values of the first time series variable and the second time series variable along the time series variable axis on a data plane formed by the time axis and the time series variable axis. And generating the first matrix by extracting a measurement value corresponding to a sliding window on the data plane.

In one embodiment, the prediction model is further based on a recurrent neural network, and predicting a class of the prediction target comprises: extracting a feature map by inputting the first matrix to the convolutional neural network. And inputting the extracted feature map into the cyclic neural network and predicting a class of the prediction target based on an output result of the cyclic neural network.

1 is a configuration diagram showing a time series data analysis system according to an embodiment of the present invention.
2 is an illustration of data sources and multiple time series data that may be referenced in some embodiments of the invention.
3 and 4 are block diagrams showing an apparatus for analyzing time series data according to an embodiment of the present invention.
5 is a hardware configuration diagram illustrating a time series data analysis device according to an embodiment of the present invention.
6 is a flowchart illustrating a method for analyzing time series data according to a first embodiment of the present invention.
7 to 9 are exemplary views for explaining a pre-processing method according to some embodiments of the present invention.
10 to 12 are exemplary views for explaining a method of generating a matrix based on multiple time series data according to an embodiment of the present invention.
13 is an exemplary view for explaining a method for determining a main influence variable according to a first embodiment of the present invention.
14 and 15 are exemplary views for explaining a method for more effectively determining a main influence variable according to an embodiment of the present invention.
16 is a flowchart illustrating a method for analyzing time series data according to a second embodiment of the present invention.
17 is an exemplary diagram for explaining the structure of a prediction model according to an embodiment of the present invention.
18 is an exemplary view for explaining a method for determining a main influence variable according to a second embodiment of the present invention.
19 is a configuration diagram illustrating an abnormality detection system according to an exemplary embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Advantages and features of the present invention, and methods for achieving them will be clarified with reference to embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only the embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention pertains. It is provided to completely inform the person having the scope of the invention, and the present invention is only defined by the scope of the claims.

It should be noted that in adding reference numerals to the components of each drawing, the same components have the same reference numerals as possible even though they are displayed on different drawings. In addition, in describing the present invention, when it is determined that detailed descriptions of related well-known structures or functions may obscure the subject matter of the present invention, detailed descriptions thereof will be omitted.

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used in a sense that can be commonly understood by those skilled in the art to which the present invention pertains. In addition, terms defined in the commonly used dictionary are not ideally or excessively interpreted unless specifically defined. The terminology used herein is for describing the embodiments and is not intended to limit the present invention. In the present specification, the singular form also includes the plural form unless otherwise specified in the phrase.

In addition, in describing the components of the present invention, terms such as first, second, A, B, (a), and (b) may be used. These terms are only for distinguishing the component from other components, and the nature, order, or order of the component is not limited by the term. When a component is described as being "connected", "coupled" or "connected" to another component, the component may be directly connected to or connected to the other component, but another component between each component It will be understood that elements may be "connected", "coupled" or "connected".

As used herein, "comprises" and / or "comprising" refers to the components, steps, operations and / or elements mentioned above, the presence of one or more other components, steps, operations and / or elements Or do not exclude additions.

Prior to the description of the present specification, some terms used in the specification will be clarified.

In this specification, multiple times series data means data composed of measured values related to two or more time series variables. The term multi-time series data may be used interchangeably with terms such as multi-dimensional time series data or multivariate time series data in the art.

In this specification, the time series variable (times series variable) refers to all variables having characteristics that can be measured or observed over time. In this case, the variable may be used interchangeably with terms such as attribute, variable, and factor in the art. Examples of the time series variable may be temperature, humidity, stock index, exchange rate, etc., but the technical scope of the present invention is not limited to the examples listed above.

In the present specification, a target of prediction literally refers to an object to be predicted by analyzing multiple time series data. For example, when a process anomaly is predicted using multiple time series data composed of measured values such as temperature and humidity, the predicted object may refer to a process state. For another example, when predicting the value of a specific item (eg company, real estate) using multi-series data composed of observed values such as an exchange rate and a composite stock price index, the prediction target refers to the value of the specific item Can be. When predicting a class (e.g. or higher, normal) of a prediction target using measurement values of a plurality of time series variables, the time series variable may correspond to an independent variable and the prediction target may correspond to a dependent variable.

In the present specification, a prediction model means a model used to predict a class of prediction targets. For example, the prediction model may be a model built through machine learning, but the technical scope of the present invention is not limited thereto.

In the present specification, instructions (instructions) refers to a set of instructions grouped by function, which are components of a computer program and executed by a processor.

Hereinafter, some embodiments of the present invention will be described in detail according to the accompanying drawings.

1 is a configuration diagram showing a time series data analysis system according to an embodiment of the present invention.

As shown in FIG. 1, the time series data analysis system may include at least one data source 10-1 to 10-n, a collection device 50, and a time series data analysis device 100. However, this is only a preferred embodiment for achieving the object of the present invention, and of course, some components may be added or deleted as necessary. In addition, it is noted that each component of the time series data analysis system shown in FIG. 1 represents functionally divided functional elements, and a plurality of components may be implemented in an integrated form in an actual physical environment. For example, the collection device 50 and the time series data analysis device 100 may be implemented in the form of different logic in the same physical computing device.

Further, in the actual physical environment, each of the components may be implemented in a form of being divided into a plurality of detailed functional elements. For example, a first function of the time series data analysis device 100 may be implemented in a first computing device constituting a computing system, and a second function may be implemented in a second computing device constituting the computing system. Hereinafter, each of the components will be described.

In the time series data analysis system, at least one data source (10-1 to 10-n) is a device or storage providing time series data to be analyzed. For example, as illustrated in FIG. 2, when the data to be analyzed is measured values related to temperature, humidity, and the like, the data sources 10-1 to 10-n may include various sensors 20-1 to provide the measured values. 20-n). For another example, when the data to be analyzed is financial data such as exchange rates and stock indices, the data sources 10-1 to 10-n may refer to a storage or device that provides the financial data.

In the time series data analysis system, the collection device 50 is a device that collects multiple time series data from at least one data source 10-1 to 10-n. For example, the collection device 50 may collect first time series data from the first data source 10-1 and second time series data from the second data source 10-2. The method in which the collection device 50 collects multiple time series data may be any method.

In the time series data analysis system, the time series data analysis device 100 is a computing device equipped with an analysis function for multiple time series data. Here, the computing device may be a laptop, a desktop, a laptop, etc., but is not limited thereto, and may include all types of devices equipped with computing functions. However, in an environment that analyzes a large amount of multi-time series data, the time series data analysis apparatus 100 may be preferably implemented as a high-performance server-class computing device. For convenience of description, hereinafter, the time series data analysis device 100 will be abbreviated as the analysis device 100.

According to an embodiment of the present invention, the analysis device 100 may provide class information for a prediction target by analyzing multiple time series data in consideration of a correlation between time series variables. For example, when the prediction target is a process state, class information (e.g. abnormality, normal) for the process state may be provided by analyzing multiple time series data in consideration of a correlation between process variables. According to the present embodiment, high-quality prediction information with high reliability can be provided by considering a correlation between time-series variables. Detailed description of this embodiment will be described later with reference to the drawings of FIG. 3 and below.

Further, according to an embodiment of the present invention, the analysis device 100 may determine a main influence variable that has the most influence on class determination among a plurality of time series variables. For example, when a process abnormality is predicted, the analysis device 100 may determine a main influence variable (ie, a main influence factor) that most affects the process abnormality among a plurality of time series variables (e.g. temperature, humidity, etc.). According to the present embodiment, cause information (ie, main influence variable) for the prediction result is additionally provided along with the prediction result (ie, class information). Therefore, there is an advantage in that high utilization and valuable information are provided. Detailed description of the present embodiment will also be described later with reference to the drawings of FIG. 3 and below.

At least some components of the time series data analysis system shown in FIG. 1 can communicate over a network. Here, the network is a wired / wireless network of any kind, such as a local area network (LAN), a wide area network (WAN), a mobile radio communication network, a Wibro (Wireless Broadband Internet), and the like. Can be implemented.

So far, a time series data analysis system according to an embodiment of the present invention has been described with reference to FIGS. 1 and 2. Hereinafter, the configuration and operation of the analysis device 100 will be described in more detail with reference to FIGS. 3 to 5.

3 is a block diagram showing an analysis device 100 according to an embodiment of the present invention.

Referring to FIG. 3, the analysis device 100 includes a data collection unit 110, a pre-processing unit 120, a matrix generation unit 130, an analysis unit 140, a pattern DB 150, and a main influence variable determination unit ( 160). However, only components related to the embodiment of the present invention are illustrated in FIG. 3. Accordingly, it can be seen that a person skilled in the art to which the present invention pertains may further include other general-purpose components in addition to the components shown in FIG. 3. In addition, it is noted that each of the components of the analysis device 100 illustrated in FIG. 3 is functionally divided functional elements, and a plurality of components may be implemented in an integrated form in a physical environment.

Looking at each component, the data collection unit 110 collects multiple time series data from at least one data source (e.g. 10-1 to 100-n). Alternatively, the data collection unit 110 may collect multiple time series data from another collection device (e.g. 50 of FIG. 1).

Next, the pre-processing unit 120 performs pre-processing on the collected multiple time series data. In order to exclude the duplicated description, a detailed description of the operation of the pre-processing unit 120 will be described later with reference to FIGS. 6 to 9.

Next, the matrix generator 130 generates a matrix from pre-processed multi-time series data. Specifically, the matrix generator 130 may extract data for each preset time series section from the pre-processed multi-time series data, and generate a matrix based on the extracted data. The generated matrix may be stored in the pattern DB 150 for each pattern. In order to exclude duplicate description, a detailed description of the operation of the matrix generator 130 will be described later with reference to FIGS. 6 and 10 to 12.

Next, the analysis unit 140 analyzes the generated matrix to predict a class of prediction targets. As shown in FIG. 4, the analysis unit 140 may include a first analysis unit 141 and a second analysis unit 143.

The first analysis unit 141 predicts the class of the prediction target based on the frequency of occurrence of the pattern matched with the generated matrix. At this time, one matrix may match one pattern, and a plurality of matrices may match one pattern. A detailed description of the operation of the first analysis unit 141 will be described later with reference to FIGS. 6 to 12.

Next, the second analysis unit 143 applies the generated matrix to the prediction model to predict the class of the prediction target. As described above, the prediction model may be a model built through machine learning. However, the scope of the present invention is not limited to this. A detailed description of the operation of the second analysis unit 143 will be described later with reference to FIGS. 16 and 17.

Referring to FIG. 3 again, the pattern DB 150 is a storage in which the matrix generated by the matrix generator 130 is stored for each pattern. The pattern DB 150 may update the frequency of occurrence of the matched pattern when the matrix is stored.

Next, the main influence variable determining unit 160 determines a main influence variable that has the most influence on class determination among a plurality of time series variables. In addition, the main influence variable determination unit 160 may determine the degree of influence on class determination for each time series variable. As illustrated in FIG. 4, the main influence variable determining unit 160 may include a first main influence variable determining unit 161 and a second main influence variable determining unit 163.

The first main influence variable determining unit 161 determines the main influence variable based on the matrix similarity. A detailed description of the operation of the first main influence variable determination unit 161 will be described later with reference to FIGS. 6 and 13 to 15.

Next, the second main influence variable determining unit 163 determines the main influence variable based on the confidence score of the prediction model. A detailed description of the operation of the second main influence variable determination unit 163 will be described later with reference to FIGS. 16 and 18.

Meanwhile, according to another embodiment of the present invention, the analysis device 100 may be implemented in a form in which some of the components illustrated in FIG. 3 are omitted. That is, it should be noted that the form shown in FIG. 3 is not the only configuration of the analysis device 100.

Each component shown in FIGS. 3 and 4 may mean software or hardware such as a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). However, the above components are not limited to software or hardware, and may be configured to be in an addressable storage medium, or may be configured to execute one or more processors. The functions provided in the above components may be implemented by more detailed components, or may be implemented as a single component that performs a specific function by combining a plurality of components.

5 is a hardware configuration diagram illustrating an analysis device 100 according to an embodiment of the present invention.

Referring to FIG. 5, the analysis device 100 includes one or more processors 101, a bus 105, a communication interface 107, and a memory 103 for loading a computer program performed by the processor 101 And a storage 109 for storing the computer program 109a. However, only components related to the embodiment of the present invention are illustrated in FIG. 5. Accordingly, it can be seen that a person skilled in the art to which the present invention belongs may include other general-purpose components other than those shown in FIG. 5.

The processor 101 controls the overall operation of each component of the analysis device 100. The processor 101 comprises a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit) or any type of processor well known in the art of the present invention. Can be. Also, the processor 101 may perform operations on at least one application or program for executing the method according to the embodiments of the present invention. The analysis device 100 may include one or more processors.

The memory 103 stores various data, commands and / or information. The memory 103 may load one or more programs 109a from the storage 109 in order to execute the time series data analysis method according to embodiments of the present invention. The memory 103 may be implemented as a volatile memory such as RAM, but the scope of the present invention is not limited thereto. When the computer program 109a is loaded into the memory 103, the module illustrated in FIG. 3 on the memory 103 may be implemented in the form of logic.

The bus 105 provides a communication function between components of the analysis device 100. The bus 105 may be implemented as various types of buses such as an address bus, a data bus, and a control bus.

The communication interface 107 supports wired and wireless Internet communication of the analysis device 100. Also, the communication interface 107 may support various communication methods other than Internet communication. To this end, the communication interface 107 may include a communication module well known in the technical field of the present invention.

The storage 109 may store multiple time series data (not shown) and the one or more programs 109a non-temporarily. The storage 109 is a non-volatile memory such as a read only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EPMROM), a flash memory, a hard disk, a removable disk, or well in the technical field to which the present invention pertains. And any known form of computer-readable recording media.

Computer program 109a may include one or more instructions that, when loaded into memory 103, cause processor 101 to perform methods in accordance with some embodiments of the present invention. The processor 101 can perform the above methods by executing the one or more instructions.

For example, the computer program 109a extracts data of a predetermined time series section from multiple time series data, generates a matrix, analyzes the generated matrix, and performs instructions for predicting a class of the prediction target. It can contain.

So far, the configuration and operation of the analysis apparatus 100 according to an embodiment of the present invention have been described with reference to FIGS. 3 to 5. Hereinafter, a method of analyzing time series data according to some embodiments of the present invention will be described in detail with reference to the drawings of FIG. 6 and below.

Each step of the time series data analysis method may be performed by a computing device. In other words, each step of the time series data analysis method may be implemented with one or more instructions executed by a processor of a computing device. All steps included in the time series data analysis method may be executed by one physical computing device, but the first steps of the method are performed by the first computing device, and the second steps of the method are second computing It can also be performed by the device. Hereinafter, it is assumed that each step of the time series data analysis method is performed by the analysis device 100 to continue the description. However, for convenience of description, description of the operation subject of each step included in the time series data analysis method may be omitted.

First, a method for analyzing time series data according to a first embodiment of the present invention will be described with reference to FIGS. 6 to 15.

6 is a flowchart illustrating a method for analyzing time series data according to a first embodiment of the present invention. However, this is only a preferred embodiment for achieving the object of the present invention, and of course, some steps may be added or deleted as necessary.

As illustrated in FIG. 6, the first embodiment starts at step S10 in which the analysis device 100 collects multiple time series data. As described above, the multiple time series data refers to data composed of measurement values for a plurality of time series variables. The method of collecting multiple time series data may be any method.

In step S30, the analysis device 100 performs pre-processing on the collected multiple time series data. The specific operation of the pre-processing step (S30) may vary depending on the embodiment.

In one embodiment, the analysis device 100 may perform data compression processing. For example, as illustrated in FIG. 7, the analysis device 100 divides the time series data 210 into preset sections (eg, first section, second section, etc.), and averages the time series data 210 for each section. Values 201 to 209 can be calculated. The graph shown at the top of FIG. 7 is time series data 210 before compression, and the graph shown at the bottom of FIG. 7 is time series data 201 to 209 after compression. According to the present embodiment, since only the average values 201 to 215 for each section are stored, not all of the time series data 210, an effect of efficiently using storage space and reducing computing cost for analysis can be achieved. . In addition, since the influence of noise is reduced through averaging, a noise reduction effect can also be achieved.

In one embodiment, the analysis device 100 may perform normalization. For example, as shown in FIG. 8, the analysis device 100 may normalize the multi-time series data 211-1 and 213-1 to values within a predetermined range using average and variance as shown in FIG. 8. have. Of course, normalization may be performed in any other way without using the mean and variance. In FIG. 8, the graph shown at the top is time series data 211-1, 213-1 before normalization, and the graph shown at the bottom is time series data 211-2, 213-2 after normalization.

In one embodiment, the analysis device 100 may perform symbolization processing. For example, as illustrated in FIG. 9, the analysis device 100 may symbolize the time series data 221 through symbolic aggregate approXimation (SAX) transformation. More specifically, the analysis device 100 fragments the time series data 221 for each section through a piecewise aggregate approximation (PAA) transformation, and a symbol matching the fragmented time series data 223 based on threshold values (eg a, b, c). 9 shows that the time series data 221 is converted to "baabccbc" as a result of symbolization, but the symbolization may also include converting the time series data to numbers other than letters such as alphabets.

In one embodiment, the analysis device 100 may perform dimensional reduction processing on multiple time series data through Principal Component Analysis (PCA). For example, the analysis device 100 may extract principal component variables of k (where k is a natural number less than n) from n time series variables through principal component analysis. Accordingly, the n-dimensional data can be reduced to k-dimensional data while maintaining the data-specific characteristics as much as possible. Since the principal component analysis is a technique well known in the art, detailed description thereof will be omitted. According to this embodiment, by reducing the dimension of the data, computing cost for analyzing time-series data can be greatly reduced.

In one embodiment, the analysis device 100 may perform a pre-processing step S30 based on a combination of the above-described embodiments.

Referring back to FIG. 6, in step S50, the analysis device 100 extracts data corresponding to a preset time series section from pre-processed multi-time series data to generate a matrix having a two-dimensional data structure. At this time, the length of the time-series section may be varied depending on the embodiment. For reference, the two-dimensional matrix only means that it can be expressed in a two-dimensional array as shown in FIGS. 10 to 12, but does not mean that the number of time series variables is two. Further, according to an embodiment, a multidimensional matrix of two or more dimensions may be generated.

In this step S50, the analysis apparatus 100 may generate a matrix for each time series section. In order to provide a more convenient understanding, this step (S50) will be described in detail with reference to FIGS. 10 to 12.

As shown in FIG. 10, the multiple time series data 231 to 236 may be disposed on a two-dimensional data plane formed by the time axis and the time series variable axis. At this time, the arrangement order of the time series variables may vary depending on the embodiment.

In one embodiment, the arrangement order may be randomly determined.

In one embodiment, the arrangement order may be determined based on a correlation analysis result between time series variables. For example, the analysis device 100 performs correlation analysis on the first time series variable and the second time series variable, and in response to determining that a correlation exists, the first time series variable and the second time series on the time series variable axis Variables can be placed adjacent to each other. If prior knowledge of the correlation is given, an order of arrangement may be determined based on the prior knowledge without performing correlation analysis. According to the present embodiment, time series variables that are most likely to be related through correlation analysis are disposed adjacent to the data plane. Therefore, the correlation between time series variables can be better reflected in the data analysis process. For example, when data analysis is performed through a convolutional neural network (CNN), the local characteristics will be better extracted, and thus the accuracy of prediction may be improved.

In one embodiment, the arrangement order of time series variables may be determined by a combination of the above-described embodiments. For example, the analysis device 100 may arrange the first plurality of time-series variables having a correlation adjacently and randomly arrange the second plurality of time-series variables having no correlation.

FIG. 11 shows a two-dimensional data plane shown in FIG. 10 in a matrix form.

As shown in FIG. 11, the first row of the matrix shown in FIG. 11 corresponds to the first time series data 231, and the second row corresponds to the second time series data 232. In addition, the first column of the matrix corresponds to the value of each time series data 231 to 236 measured at the first time point, and the second column corresponds to the value of each time series data 231 to 236 measured at the second time point. Correspond. Of course, depending on the embodiment, the correspondence between rows and columns may be changed.

12 illustrates a process of generating a matrix corresponding to each time series section in the data plane shown in FIG. 11.

As illustrated in FIG. 12, the analysis device 100 may continuously generate matrices 241-2 and 243-3 in a sliding window manner. Specifically, the analysis device 100 extracts the region 241-1 corresponding to the window set on the data plane to generate the first matrix 241-2, and the region 243-1 corresponding to the sliding window. The second matrix 243-3 may be generated by extracting. At this time, the moving interval of the window (ie, stride) and the size of the window (ie, the length of the time series section) may vary depending on the embodiment.

For example, if strict monitoring is required, such as a semiconductor process, or when a time-series relationship of data is important, the movement interval may be set to a relatively small value. By doing so, more thorough monitoring can be performed. For another example, if the computing resource of the analysis device 100 is a poor environment, the movement interval may be set to a relatively large value. This is because redundant data between the two matrices is minimized and the number of matrices to be analyzed can be reduced.

The matrices 241-2 and 243-2 generated as described above may be stored in the pattern DB 150. The pattern DB 150 may store the matrix 241-2 and 243-2, and at the same time, determine a pattern matching the matrix 241-2 and 243-2, and increase the frequency of occurrence of the pattern.

In this case, the pattern may have a 1: 1 relationship with the matrix, or may have a 1: multi relationship. For example, when the pattern and the matrix are in a 1: 1 relationship, each matrix itself may be used as a pattern. When the pattern and the matrix are in a 1: multi relationship, a representative matrix generated through clustering may be used as a pattern. The representative matrix may be, for example, generated by averaging one or more matrices belonging to a cluster, but may be generated in any other way (e.g. median, mode). In this case, one pattern matches one cluster, and the frequency of occurrence of the pattern can be calculated by the number of matrices belonging to the cluster.

It should be noted that FIGS. 10 to 12 illustrate the process of generating a matrix from multiple time series data from a conceptual point of view. In actual implementation, instead of arranging time series data on the data plane, the first row of the matrix is composed of the measured values of the first time series variable of a specific time series section, and the measured values of the second time series variable of a specific time series section are used. This is because by constructing the second row of the matrix, a matrix can be created.

Referring back to FIG. 6, in step S70, the analysis apparatus 100 predicts a class of a prediction target based on the frequency of occurrence of the pattern matching the matrix.

For example, when the class of the prediction target is the abnormal class and the normal class, the analysis apparatus 100 may predict the class of the prediction target as the abnormal class in response to a determination that the frequency of occurrence of the pattern is less than a threshold value. This is because a rare pattern with a low occurrence frequency has a high probability of being close to an abnormal class. Here, the threshold value may be a preset fixed value or a variable value that fluctuates depending on the situation.

In step S90, the analysis device 100 determines a main influence variable that has the most influence on the class determination among the plurality of time series variables. For example, when the predicted class is predicted as an abnormal class, the analysis device 100 may determine a main influence variable that has the most influence on the abnormality determination among the plurality of time series variables. Hereinafter, this step (S90) will be described in detail with reference to FIGS. 13 and 15.

13 to 15 are exemplary views for explaining a process of determining a main influence variable for an abnormal class. In order to provide convenience of understanding, FIGS. 13 to 15 show an example in which a class of a prediction target is divided into normal and abnormal, but the descriptions below are applied equally even when three or more multi-classes exist. Can be. In addition, in the following drawings, a portion shown in a shadow on the matrix indicates a portion to which a blind filter is applied, and the blind filter conceptually means that the value of the corresponding portion is excluded from the similarity calculation process. A specific method of excluding the value of the corresponding part in the similarity calculation process may be to replace the value of the corresponding part with "0", or change it to an arbitrary value, but this may be implemented in any way. It will be described below with reference to FIG. 13.

As shown in FIG. 13, the analysis device 100 predicts a matrix that is predicted as an abnormal class among a plurality of matrices (253 to 257, hereinafter referred to as “normal matrices”) classified as normal (251, hereinafter “anomaly matrix”) ). In this case, the matching condition may refer to a condition in which similarities between matrices are equal to or greater than a threshold value, but the matching condition may vary depending on embodiments.

More specifically, the analysis device 100 applies a blind filter for each row (ie, for each time series variable), and then calculates the similarity between the abnormal matrix 251 and the first normal matrix 253. For example, the analysis device 100 applies a blind filter to the first rows 251-1 and 253-1 of the two matrices 251 and 253, calculates matrix similarity, and performs the same process as the last row 251-2. , 253-2). At this time, the method of calculating the similarity may be any method.

The above process can be performed in the same way for other normal matrices (e.g. second normal matrix 255, third normal matrix 257, etc.). The analysis apparatus 100 may perform the above search process until a normal matrix satisfying the matching condition is found, or may perform the above search process for all given normal matrices.

If a matching normal matrix is found, the main influence variable can be determined immediately. For example, suppose that a matching normal matrix is found when a blind filter is applied to the first row 251-1. Then, the analysis device 100 may determine the first time series variable corresponding to the first row 251 as the main influence variable. When the measured value of the first time series variable is excluded, the abnormal matrix 251 is close to the normal matrix because it means that the measured value of the first time series variable has the greatest influence on the abnormality determination.

On the other hand, there may be a case where a large number of matched normal matrices are found and two or more time series variables are determined as main influence variables. For example, a matching normal matrix is found even when a blind filter is applied to another row (eg, last row 251-2) other than the first row 251-1, and two time series variables may be determined as the main influence variable . In this case, the analysis device 100 may determine the rank of the main influence variable according to a predetermined criterion. For example, the first time series variable and the second time series variable are determined as main influence variables, and the matrix similarity (ie, the matrix similarity calculated when the first time series variable is blocked) associated with the first time series variable is the second time series variable. If it is higher than the matrix similarity associated with the time series variable, the first time series variable may be determined as the primary influence variable of the priority. For another example, if the number of normal matrices associated with a first time series variable (ie, the normal matrix matched when the first time series variable is blocked) is greater than the number of normal matrices associated with the second time series variable, the first time series The variable can be determined as the primary influence variable.

On the other hand, since the number of normal matrices will usually be very large, calculating similarity with all normal matrices is very inefficient in terms of computing cost. Therefore, it is necessary to select a normal matrix that is the object of calculating similarity according to an appropriate criterion. According to an embodiment of the present invention, a matrix for calculating similarity may be selected by applying a Locality Sensitive Hashing (LSH) algorithm, which will be described below with reference to FIGS. 14 and 15.

As shown in FIG. 14, among the matrix 261-1 to 261-n, the first matrix 261-1 is an ideal matrix, and the remaining matrices 261-2 to 261-n are normal matrices. The analysis device 100 is a signature vector (signature vector 263-1 to 263-n) or a signature matrix (hereinafter, "signature") from each matrix 261-1 to 261-n through minimum hashing. (Collectively as). The minimum hashing is a technique that preserves similarity (e.g. Jakard similarity) and transforms a large set into a signature of a small size. Since the minimum hashing is a technique well known in the art, detailed description thereof will be omitted.

Next, as shown in FIG. 15, the analysis device 100 synthesizes the generated signatures 263-1 to 263-n to generate a matrix 265. In addition, the analysis apparatus 100 applies the LSH algorithm to divide the matrix 265 into b bands, and calculates a hash value for each band. The bucket shown in FIG. 15 conceptually indicates a set of bands having the same hash value.

Here, the analysis device 100 may select a matrix to calculate the similarity of the normal matrix (eg, the matrix of signature 263-k) existing in the same bucket as the band (eg 266) constituting the first signature 263-1. have. For example, since the specific band 266 of the first signature 263-1 and the band 267 of the k-th signature 263-k exist in the same bucket, the k-th signature 263-k is normal The matrix may be selected as a matrix for calculating similarity. This is because, due to the characteristics of minimum hashing and LSH algorithms, two matrices having the same hash value of some bands are more likely to be similar to each other.

The analysis device 100 includes first normal matrices present in the same bucket as the first band and second normal matrices present in the same bucket as the second band among the b bands constituting the first signature 263-1. All can be selected as a target for calculating similarity. Alternatively, only the matrices belonging to the intersection of the first normal matrices and the second normal matrices may be selected as a similarity calculation target.

According to the above-described embodiment, since some normal matrices predicted to have high similarity are selected, and the matrix similarity operation is performed only on the selected normal matrices, computing cost required to determine the main influence variable can be greatly reduced.

Meanwhile, according to another embodiment of the present invention, the analysis device 100 may select a matrix for calculating similarity through a clustering technique. Specifically, the analysis device 100 may cluster a normal matrix set to build a predetermined number of clusters, and select a representative matrix of each cluster as a target for calculating similarity. According to the present embodiment, by selecting only a representative pattern for the normal matrix as a similarity calculation target, computing cost can be greatly reduced.

So far, with reference to FIGS. 13 to 15, a method for determining a main influence variable according to an embodiment of the present invention has been described. In the above-described method, the classes to be predicted are limited to normal and abnormal classes. However, it should be noted that the method for determining the main influence variable can be performed in the same manner when determining the main influence variable for any first class. For example, suppose that there is a first matrix predicted as a first class and at least one second matrix corresponding to a second class, and determines a main influence variable for the first class. In this case, the analysis device 100 may determine a main influence variable for the first class by applying a blind filter to the two matrices and calculating the similarity between the first matrix and the second matrix.

Meanwhile, according to an embodiment of the present invention, a matrix of the same class may be used to determine a main influence variable for the first class. For example, suppose that a matrix predicted as an anomaly class is a first anomaly matrix, and a matrix pre-classified as an anomaly class is a second anomaly matrix. Then, the analysis device 100 may apply a blind filter for each row of the first abnormal matrix and the abnormal second matrix (that is, for each time series variable), and then calculate the similarity between the two matrices. In addition, when the matrix similarity (ie, the similarity calculated when the first time series variable is blocked) associated with the first time series variable is lower than or equal to the threshold of the matrix similarity associated with another time series variable (ie, when the difference is greater than or equal to the threshold value), The first time series variable may be determined as a main influence variable for the abnormal class. When the measurement values of the first time series variable are excluded, the specific abnormal matrix is not most similar to the other abnormal matrix because it means that the measurement value of the first time series variable has the greatest influence on the abnormality determination. Of course, the conditions for determining the main influence variable may be modified in any number depending on the embodiment.

For reference, among the above-described steps (S10 to S90), step (S10) may be performed by the data collection unit 110, and step (S30) may be performed by the pre-processing unit 120. In addition, step S50 is performed by the matrix generator 130, step S70 is performed by the first analysis unit 141, and step S90 is the first main influence variable determination unit 161 Can be performed by.

So far, the method of analyzing time series data according to the first embodiment of the present invention has been described with reference to FIGS. 6 to 15. According to the method described above, multiple time series data are processed into a matrix of two-dimensional data structures. The two-dimensional data structure is a data structure suitable for reflecting autocorrelation of time series data and correlation between time series variables. Therefore, in analyzing multiple time series data and performing prediction, accuracy of analysis and prediction can be greatly improved. Furthermore, a main influence variable that influences class determination can be accurately identified by using a blind filter.

Hereinafter, a method of analyzing time series data according to a second embodiment of the present invention will be described with reference to FIGS. 16 to 18.

16 is a flowchart illustrating a method for analyzing time series data according to a second embodiment of the present invention. In the following description, items that overlap with the contents of the first embodiment mentioned above will be omitted for clarity of specification.

16, the overall process of the second embodiment is similar to the first embodiment described above. However, the second embodiment differs from the first embodiment described in step S170 and S190 in that it predicts a class of a prediction target based on a prediction model and determines a main influence variable.

The prediction model is a model used to predict a class of prediction targets. The predictive model may be constructed through machine learning, but the specific configuration and operation method of the predictive model may vary according to embodiments.

In one embodiment, the prediction model may be constructed based on a convolutional neural network. The convolutional neural network is a specialized neural network for extracting regional features from data of two or more dimensions. Therefore, the convolutional neural network is the most suitable model for extracting features by considering the correlation between time series variables in matrix data. In some embodiments, before the matrix is input to the convolutional neural network, a process of appropriately correcting the matrix value to a range of pixel values may be performed. Of course, the correction process may be performed together with other pre-processing steps in step S130. The convolutional neural network is a neural network specialized for an image classification task, and those skilled in the art will be able to clearly understand the configuration and operation of the convolutional neural network, and detailed description thereof will be omitted.

In this embodiment, the analysis device 100 may predict a class of a prediction target based on a confidence score for each class output by the prediction model. For example, the analysis apparatus 100 may predict the class of the prediction target as the first class in response to the determination that the confidence score of the first class is the highest. According to this embodiment, accurate analysis and prediction can be performed by utilizing the characteristics of a convolutional neural network.

In one embodiment, the prediction model may be constructed based on a combination of a convolutional neural network and a recurrent neural network (RNN). The cyclic neural network is a neural network specialized for extracting features over time through a cyclic connection structure. In addition, time series data generally have an auto-correlation, so that the data of the past affects the current data. Therefore, when the two neural networks are combined, characteristics of time series data having autocorrelation can be better considered, and accurate analysis of multiple time series data can be achieved.

For a more specific example, the prediction model may be configured based on a convolutional neural network 273 and a long short-term memory model (LSTM) neural network 275, which is a type of cyclic neural network. . In this case, the convolutional neural network 273 receives the multiple matrices 271-1 to 271-n generated in step 150, and features (eg, from the multiple matrices 271-1 to 271-n) Feature map). Also, the LSTM neural network 275 performs an operation of outputting a confidence score for each class of a prediction target based on the features extracted from the convolutional neural network 273. As in the previous embodiment, the analysis device 100 may predict the class of the prediction target based on the confidence score for each class.

According to this embodiment, characteristics of time-series data having autocorrelation through a combination of a convolutional neural network and a circulating neural network can be considered in depth. Accordingly, the accuracy of analysis and prediction can be further improved.

Next, a method of determining the main influence variable in step 190 will be described in detail with reference to FIG. 18.

18 is an exemplary view for explaining a method of determining a main influence variable using a prediction model according to an embodiment of the present invention. In order to provide convenience for understanding, FIG. 18 also shows a case in which a class to be predicted is divided into normal and abnormal, but the following description may be applied equally even when there are three or more multiple classes. Hereinafter, it will be described with reference to FIG. 18.

As shown in FIG. 18, the analysis device 100 obtains the confidence scores 283 and 285 for each class by applying the anomaly matrix 281 to which a blind filter is applied to a specific row (ie, a time series variable) to a prediction model. Can be. For example, the analysis device 100 applies a blind filter to the first row 281-1 of the anomaly matrix 281 to calculate a confidence score 283, and repeats this process to the last row 281-2. Can be.

Here, the analysis device 100 may determine a specific time series variable that satisfies a predetermined condition as a main influence variable, a confidence score for each associated class among a plurality of time series variables (ie, a confidence score calculated when the corresponding time series variable is blocked). have.

In this case, the predetermined condition is a first condition indicating when a confidence class of the normal class associated with the specific time series variable (hereinafter, “normal confidence score”) is equal to or greater than a threshold value, and the normal confidence score associated with the specific time series variable is originally ( That is, a normal condition score associated with another time series variable or a normal confidence score associated with the specific time series variable or a second condition indicating when the threshold filter is higher than a threshold value (i.e., when the difference is greater than or equal to a threshold) than when a blind filter is not applied at all. A third condition higher than a threshold value may be included.

Alternatively, the predetermined condition is a fourth condition indicating when the confidence score of the abnormal class associated with the specific time series variable (hereinafter, “ideal confidence score”) is less than a threshold value, and the abnormal confidence score associated with the specific time series variable is originally ( In other words, a fifth condition indicating when the threshold value is lower than a threshold value (i.e., when the difference is greater than or equal to a threshold) than when a blind filter is not applied at all, or an abnormality confidence score associated with another time series variable having an abnormality confidence score associated with the specific time series variable A sixth condition indicating a case lower than or equal to a threshold value may be included. However, since the examples of the conditions listed above are for explaining some embodiments of the present invention, the technical scope of the present invention is not limited to the examples listed above.

For reference, among the aforementioned steps (S170), step 170 may be performed by the second analysis unit 143, and step S190 may be performed by the second main influence variable determination unit 163.

So far, the method of analyzing time series data for the second embodiment of the present invention has been described with reference to FIGS. 16 to 18. According to the above-described method, analysis and prediction of multiple time series data may be performed through a convolutional neural network based prediction model. The convolutional neural network is a machine learning model specialized for extracting regional features from data of two or more dimensions. Therefore, the correlation between time series variables can be well reflected in the analysis and prediction process, and accordingly, the accuracy of analysis and prediction on time series data can be greatly improved. Furthermore, unlike the first embodiment, by using a predictive model and a blind filter, the main influence variable can be identified in a simple manner without searching for a similar matrix.

Hereinafter, an example in which the technical idea of the present invention is utilized in the field of process anomaly detection will be briefly described with reference to FIG. 19 to provide more convenience.

19 is a configuration diagram showing a process anomaly detection system according to an embodiment of the present invention.

19, the above-described technical idea of the present invention may be embodied through the anomaly detection device 300.

The abnormality detection device 300 is a device that detects an abnormality of the process facility 310 in real time by analyzing multiple time series data composed of measurement values of a plurality of sensors 320-1 to 320-n.

The anomaly detection apparatus 300 may generate a matrix from the multiple time series data and analyze the matrix to determine whether the process facility 310 is abnormal. For example, as in the above-described first embodiment (eg, see FIG. 6), the anomaly detection apparatus 300 responds to the determination that the frequency of occurrence of the pattern matching the matrix is less than the threshold, and is abnormal in the process facility 310. It can be judged that there is. For another example, as in the second embodiment described above (see e.g. FIG. 16), the anomaly detection device 300 may determine whether the process facility 310 is abnormal based on the confidence score of the predictive model. Also, in response to the abnormality determination, the abnormality detection device 300 may provide a predetermined alarm to the administrator. Through this, not only the efficiency of the manufacturing process is improved, but also the effect of increasing the convenience of management can be achieved.

Furthermore, the anomaly detection device 300 may determine a main influence variable that has the most influence on the anomaly determination among a plurality of time series variables (ie, sensor measurement variables). Through this, since information on the cause of the abnormality can be additionally provided to the manager, the efficiency of the manufacturing process and the convenience of management can be further increased.

Meanwhile, according to another embodiment of the present invention, the anomaly detection apparatus 300 may perform anomaly detection based on the combination of the first and second embodiments described above. Specifically, the anomaly detection device 300 performs anomaly detection according to the first embodiment, accumulates a training dataset, trains a predictive model using the accumulated training dataset, and then the second embodiment. Accordingly, anomaly detection can be performed based on a predictive model.

Incidentally, by performing anomaly detection according to the first embodiment, a labeling operation of assigning a class label to the matrix based on the frequency of occurrence of the pattern may be performed. The anomaly matrix and the normal matrix on which the labeling operation has been performed can be used as a training dataset of the predictive model. That is, the anomaly detection device 300 may train a predictive model with the training dataset. When the prediction model is sufficiently trained, anomaly detection may be performed according to the second embodiment.

In some embodiments, the anomaly detection apparatus 100 may perform the first anomaly detection according to the first embodiment and the second anomaly detection according to the second embodiment. In this case, the anomaly detection apparatus 100 may adjust the utilization ratio of the first anomaly detection process and the second anomaly detection in proportion to the learning maturity of the predictive model. For example, the anomaly detection apparatus 100 may increase the utilization ratio of the second anomaly detection as the learning maturity increases, and decrease the utilization ratio of the first anomaly detection. This is because the accuracy of the model will improve as the learning maturity of the machine learning model increases. According to the present exemplary embodiment, an effect of gradually improving the accuracy of anomaly detection can be achieved by increasing the utilization weight of the predictive model over time.

So far, with reference to FIG. 19, an example in which the technical idea of the present invention has been applied to the process anomaly detection field has been briefly described. As described above, by performing anomaly detection in consideration of a correlation between time-series variables, a conventional problem in which anomaly detection accuracy is poor can be solved.

On the other hand, it should be noted that the technical idea of the present invention can be applied not only to the process anomaly detection field but also to various fields dealing with multiple time series data. For example, even in the case of predicting the fluctuation of the value of a specific asset (e.g. stock, real estate, etc.) by analyzing multiple time series data on exchange rates, stock indices, etc., the above-described technical ideas of the present invention can be applied without any substantial change. Furthermore, according to the embodiments of the present invention, it is possible to accurately identify up to which of the influencing factors such as exchange rates and stock indices is the main influencing factor having the greatest influence on the fluctuation of the value.

So far, some embodiments of the present invention and effects according to the embodiments have been described with reference to FIGS. 1 to 19. The effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

The concept of the present invention described so far with reference to FIGS. 1 to 19 may be embodied as computer readable code on a computer readable medium. The computer-readable recording medium may be, for example, a removable recording medium (CD, DVD, Blu-ray Disc, USB storage device, removable hard disk), or a fixed recording medium (ROM, RAM, computer-equipped hard disk). Can be. The computer program recorded on the computer-readable recording medium may be transmitted to another computing device through a network such as the Internet and installed on the other computing device, and thus used on the other computing device.

In the above, even if all the components constituting the embodiments of the present invention are described as being combined or operated as one, the present invention is not necessarily limited to these embodiments. That is, within the object scope of the present invention, all of the components may be selectively combined and operated.

Although the operations in the drawings are shown in a specific order, it should not be understood that the operations must be performed in a specific order or in a sequential order, or that all the illustrated actions must be executed to obtain a desired result. In certain situations, multitasking and parallel processing may be advantageous. Moreover, the separation of various configurations in the above-described embodiments should not be understood as such separation is necessary, and the described program components and systems may generally be integrated together into a single software product or packaged into multiple software products. It should be understood that there is.

Although the embodiments of the present invention have been described with reference to the accompanying drawings, a person skilled in the art to which the present invention pertains may be implemented in other specific forms without changing the technical concept or essential features of the present invention. Can understand. Therefore, it should be understood that the above-described embodiments are illustrative in all respects and not restrictive. The scope of protection of the present invention should be interpreted by the claims below, and all technical spirits within the equivalent range should be interpreted as being included in the scope of the present invention.

Claims (18)

A method for determining a main influence variable for a target class among a plurality of time series variables in a computing device,
Obtaining a first matrix predicted by the target class from multiple time series data associated with the plurality of time series variables;
Obtaining a second matrix belonging to a specific class;
Calculating similarities between the two matrices except for the values of the first time series variables in the first matrix and the second matrix; And
In response to the determination that the calculated similarity satisfies a predetermined condition, comprising the step of determining the first time series variable as the main influence variable,
The first row or first column of the first matrix is composed of measured values of the first time series variable,
The second row or the second column of the first matrix is characterized by consisting of the measured value of the second time series variable,
How to determine the main impact variable. According to claim 1,
The specific class is a class different from the target class,
Determining as the main impact variable,
And in response to a determination that the calculated similarity is equal to or greater than a threshold value, determining the first time series variable as the main influence variable.
How to determine the main impact variable. According to claim 1,
The specific class is the same class as the target class,
Determining as the main impact variable,
And in response to the determination that the calculated similarity is less than a threshold value, determining the first time series variable as the main influence variable.
How to determine the main impact variable. According to claim 1,
The obtaining of the first matrix may include:
Generating the first matrix by extracting data of a preset time series section from the multiple time series data; And
And predicting a class of the first matrix as the target class based on an analysis result of the first matrix,
How to determine the main impact variable. According to claim 4,
The target class is an abnormal class,
The predicting by the target class,
And in response to a determination that the frequency of occurrence of the first pattern matching the first matrix is below a threshold, predicting the class of the first matrix as the abnormal class,
How to determine the main impact variable. According to claim 4,
The step of generating the first matrix may include:
Normalizing the multiple time series data; And
And generating the first matrix based on the normalized multiple time series data.
How to determine the main impact variable. The method of claim 6,
Generating the first matrix based on the normalized multiple time series data,
Symbolizing the normalized multiple time series data through symbolic aggregate approximation (SAX) transformation; And
And generating the first matrix based on the symbolized multi-time series data.
How to determine the main impact variable. According to claim 4,
The step of generating the first matrix may include:
Arranging measurement values for the first time series variable and the second time series variable along the time series variable axis on a data plane formed by the time axis and the time series variable axis; And
Characterized in that it comprises the step of generating the first matrix by extracting a measurement value corresponding to a sliding window (sliding window) on the data plane,
How to determine the main impact variable. According to claim 4,
The predicting by the target class,
Inputting the first matrix into a prediction model composed of a convolutional neural network; And
And predicting a class of the first matrix as the target class based on an output result of the prediction model.
How to determine the main impact variable. According to claim 1,
The obtaining of the second matrix may include:
Obtaining a plurality of candidate matrices belonging to the specific class; And
And selecting a second matrix from the plurality of candidate matrices by applying a Locality Sensitive Hashing (LSH) algorithm,
How to determine the main impact variable. A method of determining a main influence variable for a specific class among a plurality of time series variables in a computing device,
Generating a first matrix by extracting data of a preset time series section from multiple time series data associated with the plurality of time series variables;
Inputting the first matrix into a prediction model, and predicting a class of the first matrix as a first class based on a first confidence score output from the prediction model; And
Determining a main influence variable for the first class,
The first row or first column of the first matrix is composed of measured values of the first time series variable,
The second row or second column of the first matrix consists of measured values of the second time series variable,
Determining the main influence variable,
Re-entering the first matrix from which the value of the first time series variable is excluded, into the prediction model to obtain a second confidence score; And
And in response to determining that the second confidence score satisfies a predetermined condition, determining the first time-series variable as a primary influence variable of the first class.
How to determine the main impact variable. The method of claim 11,
The first confidence score is a confidence score for the first class,
The second confidence score is a confidence score for the second class,
Determining the first time series variable as the primary influence variable of the first class,
And in response to determining that the second confidence score is equal to or greater than a threshold value, determining the first time series variable as a primary influence variable of the first class.
How to determine the main impact variable. The method of claim 11,
The first confidence score and the second confidence score are both confidence scores for the first class,
Determining the first time series variable as the primary influence variable of the first class,
And in response to a determination that the difference between the first confidence score and the second confidence score satisfies a predetermined condition, determining the first time series variable as the main influence variable.
How to determine the main impact variable. In the computing device using a convolutional neural network (convolutional neural network) based prediction model in the method of analyzing the multi-time series data associated with the prediction target,
Extracting data of a preset time series section from the multiple time series data, and generating a first matrix; And
And predicting a class of the prediction target by applying the first matrix to the prediction model,
The multiple time series data includes measured values of the first time series variable and the second time series variable,
The first row or first column of the first matrix is composed of measured values of the first time series variable for the time series section,
Characterized in that the second row or the second column of the first matrix is composed of measured values of the second time series variable for the time series section,
Methods for analyzing time series data. The method of claim 14,
The step of generating the first matrix may include:
Arranging measurement values for the first time series variable and the second time series variable along the time series variable axis on a data plane formed by the time axis and the time series variable axis; And
Characterized in that it comprises the step of generating the first matrix by extracting a measurement value corresponding to a sliding window (sliding window) on the data plane,
Methods for analyzing time series data. The method of claim 15,
On the time-series variable axis, the arrangement positions of the first time-series variable and the second time-series variable are:
Characterized in that it is determined based on the correlation analysis result of the first time-series variable and the second time-series variable,
Methods for analyzing time series data. The method of claim 14,
The prediction model is further based on a recurrent neural network,
The step of predicting the class of the prediction target,
Extracting a feature map by inputting the first matrix to the convolutional neural network; And
And inputting the extracted feature map into the cyclic neural network, and predicting the class of the prediction target based on the output result of the cyclic neural network.
Methods for analyzing time series data. The method of claim 14,
The prediction target class includes a normal class and an abnormal class,
Further comprising training the predictive model,
The step of training the prediction model,
Generating a plurality of training matrices based on the collected multi-time series data;
Generating a training dataset by assigning an abnormal class to a matrix in which the frequency of occurrence of a matched pattern among the plurality of training matrices is less than a threshold and a normal class to the remaining matrices; And
Characterized in that it comprises the step of training the predictive model using the training data set,
Methods for analyzing time series data.

Images