JP2020107138A - Classification device, classification system, classification method and program - Google Patents

Abstract

To provide a classification device, a classification system, a classification method and a program capable of further improving accuracy of classification while further suppressing an increase in processing amount.SOLUTION: A classification device comprises: a processing unit acquiring non-negative evaluation data in a predetermined data area from time series data; a storage unit storing a supervised coefficient matrix generated from the non-negative first supervised data corresponding to the predetermined data area, and a classification model parameter generated based on a first class mapped to the first supervised data; a feature extraction unit generating an evaluation coefficient matrix from the evaluation data based on a supervised basic matrix generated from the first supervised data; a classification unit classifying the evaluation coefficient matrix into the first class based on the classification model parameter; and an area specification unit specifying the predetermined data area.SELECTED DRAWING: Figure 8

Description

The present invention relates to a classification device, a classification system, a classification method and a program.

Conventionally, a technique using non-negative matrix factorization has been researched in order to analyze and classify data such as acoustic signals, which is an example of continuously generated time series data such as signals obtained by sensing (see below. Patent Documents 1 to 4 and Non-Patent Documents 1 and 2 below.

JP, 2014-137389, A JP, 2005-079110, A JP, 2011-133780, A JP, 2017-151872, A

Lee, H.; S. Seung, “Learning the parts of objects by nonneatactive matrix factorization,” Nature, (UK), Oct. 1999, vol. 401 pp. 788-791. Lee, H.; S. Seung, "Algorithms for non-negative matrix factorization," Advances in neural information processing systems, Dec. 2001, pp. 556-562.

The techniques for analysis and classification as described above can analyze and classify time-series data, but in recent years, it has been strongly demanded to further improve the accuracy of analysis and classification. Further, at the same time as the above request, in order to avoid an increase in processing time and a burden on the processing device, it is required to further reduce the processing amount (processing data amount) in the above-described analysis and classification processing. Has been.

Therefore, the present invention has been made in view of the above situation, and an object of the present invention is to further suppress the increase in the processing amount and further improve the accuracy of classification, which is new and An object of the present invention is to provide an improved classification device, classification system, classification method and program.

In order to solve the above problems, according to an aspect of the present invention, there is provided a classification device for classifying time series data into an intended first class, wherein the time series represented at least in a frequency domain or a time domain. A processing unit for acquiring non-negative evaluation data of a predetermined data area from data, and a teacher generated by performing non-negative matrix factorization on the non-negative first teacher data corresponding to the predetermined data area. A storage unit for storing a coefficient matrix and a classification model parameter generated based on the first class associated with the first teacher data; and a non-negative matrix factor for the first teacher data. A feature extraction unit that generates an evaluation coefficient matrix from the evaluation data, based on the teacher basis matrix generated by performing decomposition,
The calculated evaluation coefficient matrix, based on the classification model parameter, a classification unit that classifies into the first class, and a region specifying unit that specifies the predetermined data region, the region specifying unit, Each non-negative second teacher data associated with each of the first classes is divided in the frequency domain or the time domain, and based on the observation matrix of each second teacher data for each divided area. Calculating a first evaluation index indicating the degree of separation between the first classes for each of the divided regions, and assigning each non-negative third teacher associated with each of the second classes not intended. The data is divided in the frequency domain or the time domain, and the degree of separation between the second classes for each divided area is calculated based on the observation matrix of each of the third teacher data for each divided area. The second evaluation index shown is calculated, and the first area is specified by extracting the divided area in which the first evaluation index is higher than a preset first threshold value. 2 evaluation index is higher than a preset second threshold value, the second area is specified by extracting the divided area, and based on the specified first and second areas, A classification device is provided that identifies the predetermined data area.

In order to solve the above-mentioned problems, according to another aspect of the present invention, a classification including a classification device that classifies time-series data into an intended first class and a sensor that acquires the time-series data. In the system, the classification device corresponds to the processing unit that acquires non-negative evaluation data of a predetermined data region from the time series data represented in at least the frequency domain or the time domain, and the predetermined data region. It is generated based on a teacher coefficient matrix generated by performing non-negative matrix factorization on the non-negative first teacher data, and the first class associated with the first teacher data. Feature extraction for generating an evaluation coefficient matrix from the evaluation data based on a storage unit that stores a classification model parameter and a teacher basis matrix generated by performing non-negative matrix factorization on the first teacher data A region, a classification unit that classifies the calculated evaluation coefficient matrix into the first class based on the classification model parameter, and a region specifying unit that specifies the predetermined data region. The specifying unit divides each non-negative second teacher data associated with each of the first classes in a frequency domain or a time domain, and divides each second teacher data of each of the divided areas. A first evaluation index indicating the degree of separation between the first classes for each of the divided regions is calculated based on an observation matrix, and each non-negative value associated with each of the second classes not intended The third teacher data is divided in the frequency domain or the time domain, and based on the observation matrix of each of the third teacher data for each divided area, the second class between the divided areas is divided. A second evaluation index indicating the degree of separation of the first area, and the first area is identified by extracting the divided area in which the first evaluation index is higher than a preset first threshold value. Then, the second area is specified by extracting the divided area in which the second evaluation index is higher than a preset second threshold value, and the specified first and second areas A classification system is provided for identifying the predetermined data area based on

In order to solve the above-mentioned problems, according to still another aspect of the present invention, there is provided a classification method for classifying time series data into a desired first class, which is expressed in at least a frequency domain or a time domain. It is generated by acquiring non-negative evaluation data of a predetermined data area from the time series data and performing non-negative matrix factorization on the non-negative first teacher data corresponding to the predetermined data area. A training coefficient matrix, and a classification model parameter generated based on the first class associated with the first training data, and a non-negative matrix for the first training data. Generating an evaluation coefficient matrix from the evaluation data based on a teacher basis matrix generated by performing factorization, and calculating the calculated evaluation coefficient matrix based on the classification model parameter. Classifying into a class and identifying the predetermined data area, wherein identifying the predetermined data area is performed by each non-negative second teacher associated with each of the first classes. The data is divided in the frequency domain or the time domain, and the degree of separation between the first classes of the divided areas is calculated based on the observation matrix of the second teacher data of the divided areas. The first evaluation index shown is calculated, and each non-negative third teacher data associated with each second target class is divided in the frequency domain or the time domain. A second evaluation index indicating the degree of separation between the second classes for each of the divided regions is calculated based on the observation matrix of each of the third teacher data, and the first evaluation index is set in advance. By dividing the divided area having a higher threshold value than the predetermined first threshold value to identify the first area, and the second evaluation index being higher than a preset second threshold value. A classification method is provided that includes identifying a second area by extracting the identified area and identifying the predetermined data area based on the identified first and second areas.

In order to solve the above-mentioned problems, according to still another aspect of the present invention, there is provided a program for classifying time series data into a desired first class, the program including at least a frequency domain or a time domain. A function of acquiring non-negative evaluation data of a predetermined data area from the time-series data expressed in (3) and non-negative matrix factorization of the non-negative first teacher data corresponding to the predetermined data area A teacher coefficient matrix generated by the above, and a function of storing a classification model parameter generated based on the first class associated with the first teacher data; Based on the teacher basis matrix generated by performing non-negative matrix factorization, the function of generating an evaluation coefficient matrix from the evaluation data, the calculated evaluation coefficient matrix, based on the classification model parameters, In the function of realizing the function of classifying into the first class and the function of specifying the predetermined data area and specifying the predetermined data area, a non-negative value associated with each of the first classes is provided. Of the second teacher data of each of the divided regions in the frequency domain or the time domain, and based on the observation matrix of the second teacher data of each of the divided regions, A first evaluation index indicating the degree of separation between classes is calculated, and each non-negative third teacher data associated with each of the second classes not intended is divided in the frequency domain or the time domain, A second evaluation index indicating the degree of separation between the second classes for each divided region is calculated based on the observation matrix of each of the third teacher data for each divided region, and the first evaluation index is calculated. The first threshold is specified by extracting the divided area in which the evaluation index of is higher than the first threshold set in advance, and the second threshold in which the second evaluation index is set in advance. A second area is identified by extracting the divided area that is higher than the first area, and the predetermined data area is identified based on the identified first and second areas. The program is provided.

As described above, according to the present invention, it is possible to further suppress the increase in the processing amount and further improve the classification accuracy.

It is explanatory drawing which shows an example of a schematic structure of the information processing system 10 which concerns on embodiment of this invention. It is a block diagram for explaining an example of composition of information processor 200 concerning an embodiment of the present invention. It is explanatory drawing which shows typically the example of a process of NMF which concerns on embodiment of this invention. It is a schematic diagram for demonstrating the reduction of the dimension by NMF. It is a schematic diagram which shows the example which applied NMF to the whole time/frequency/amplitude data. It is a schematic diagram which shows the example which applied NMF to the part of time/frequency/amplitude data. It is an explanatory view showing an example of data division concerning an embodiment of the present invention. It is explanatory drawing (the 1) which shows the example of acquisition of the teacher data and evaluation data which concern on embodiment of this invention. It is explanatory drawing (the 2) which shows the example of acquisition of the teacher data and evaluation data which concern on embodiment of this invention. It is a flowchart figure which shows the operation example of embodiment of this invention in a learning phase. It is a flowchart figure which shows the operation example of embodiment of this invention in a classification phase. It is an explanatory view showing a hardware configuration example.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this specification and the drawings, constituent elements having substantially the same functional configuration are designated by the same reference numerals, and a duplicate description will be omitted. In this specification and the drawings, similar components may be distinguished by attaching different alphabets after the same reference numerals. However, if there is no particular need to distinguish between similar components, only the same reference numerals will be given.

<<Background>>
First, before describing the embodiments of the present invention, the background leading to the creation of the embodiments of the present invention by the present inventor will be described.

As described above, a technique using non-negative matrix factorization to analyze and classify data such as acoustic signals, which is one example of continuously generated time series data such as signals obtained by sensing, as described above. Is being studied.

The technique as described above is an effective means for analyzing and classifying time series data. However, in recent years, as data to be analyzed and classified, even for various kinds of actually existing time series data that are not suitable data, there is a strong demand for higher accuracy of classification. .. Moreover, since various time-series data that actually exist as described above are often data in which various parameters are combined in a complicated manner, the processing time when processing such time-series data is The tendency that the number of processing devices increases and the load on the processing device also increases has become remarkable.

Therefore, in view of the above situation, the present inventor has created an embodiment of the present invention capable of further improving the classification accuracy while further suppressing an increase in the processing amount. Hereinafter, such embodiments of the present invention will be sequentially described in detail.

<<Embodiment>>
In the embodiment described below, the time-series data acquired from the sensor is obtained by, for example, collecting a tapping sound when hitting an object to be measured (such as concrete) with a hitting tool such as a hammer. It may be an acoustic signal that has been recorded. Furthermore, in the following description, as an example, the information processing device is divided into two classes (first target class) indicating whether the measurement object is normal or abnormal based on the acoustic signal. Although an example having a function of automatically classifying will be described, the present invention is not limited to such an example.

Further, in the following description, for time series data from the sensor, a power spectrum obtained by Fourier transform, a spectrogram obtained by short-time Fourier transform, a scatogram obtained by wavelet transform, etc., time, frequency, Data representing the relationship between amplitudes is generically called time/frequency/amplitude data. That is, it can be said that the time/frequency/amplitude data is data expressed at least in the frequency domain or the time domain, and can be expressed as, for example, data indicating amplitude values in the frequency domain and the time domain.

Also, in the following explanation, the data to be classified is called evaluation data, and before classifying the evaluation data, the data for making a model necessary for classification by associating with the accurate label (class) in advance is classified. Called teacher data. The correct label is a label indicating the class to which each observation vector included in the teacher observation matrix (non-negative teacher data) obtained by converting the teacher data belongs. In addition, in the following description, the basis matrix and the coefficient matrix obtained by performing non-negative matrix factorization of the teacher observation matrix obtained by converting the teacher data are referred to as a teacher basis matrix and a teacher coefficient matrix, respectively. Furthermore, in the following description, the basis matrix and coefficient matrix obtained by performing non-negative value matrix factorization of the evaluation observation matrix (non-negative evaluation data) obtained by converting the evaluation data are the evaluation basis matrix and the evaluation matrix, respectively. Called the coefficient matrix.

<Schematic configuration of information processing system 10>
First, a schematic configuration of an information processing system (classification system) 10 according to the embodiment of the present invention will be described with reference to FIG. FIG. 1 is an explanatory diagram showing an example of a schematic configuration of an information processing system 10 according to the present embodiment.

The information processing system 10 according to the present embodiment acquires, for example, time-series data (acoustic data) of tapping sound with respect to a measurement target, and based on the acquired time-series data, whether the measurement target is normal or abnormal. It is a classification system that classifies whether or not there is. As shown in FIG. 1, the information processing system 10 according to the present embodiment mainly includes a sensor 100 for acquiring the time series data and an information processing device (classification device) 200 that performs classification. The outline of each device included in the information processing system 10 will be described below.

(Sensor 100)
The sensor 100 detects a change in the state of the observation target as a physical change, and outputs the detected time series data to the information processing device 200. For example, the sensor 100 may be an acoustic sensor such as a microphone that acquires ambient sound as an acoustic signal, an acceleration sensor that acquires acceleration data, a temperature sensor that acquires temperature data, or the like. There is no particular limitation as long as it can acquire series data.

(Information processing device 200)
The information processing device 200 is a device that analyzes and classifies the time-series data output from the sensor 100. More specifically, the information processing apparatus 200 uses the time-series data (teacher data) generated continuously such as the signal obtained by the sensor 100 to generate a classification model parameter for classifying into a plurality of classes. It has a function as a device and a function as a classification device for classifying time series data (evaluation data) into a plurality of classes. For example, as shown in FIG. 1, the information processing device 200 can be a cloud 200a, an IoT gateway 200b, an edge terminal 200c, or the like. The detailed configuration of these information processing devices 200 will be described later. The information processing system 10 according to the present embodiment may include at least one information processing device 200 such as the cloud 200a, the IoT gateway 200b, the edge terminal 200c, and the like.

<Detailed Configuration of Information Processing Device 200>
The schematic configuration of the information processing system 10 according to the embodiment of the present invention has been described above. Next, a detailed configuration of an example of the information processing device 200 according to the embodiment of the present invention will be described with reference to FIG. FIG. 2 is a block diagram for explaining a configuration example of the information processing device 200 according to the embodiment of the present invention. As shown in FIG. 2, the information processing apparatus 200 according to the present embodiment has a data acquisition unit 202, an operation unit 204, a processing unit 210, a teacher data processing unit 220, a parameter generation unit 222, a storage unit 224, and a feature extraction unit 226. , A classification unit 228, and an output unit 230. Each functional unit of the information processing device 200 will be described below.

(Data acquisition unit 202)
The data acquisition unit 202 acquires the time-series data expressed in at least the frequency domain or the time domain from the sensor 100, and outputs the acquired time-series data to the processing unit 210 described below.

(Operation unit 204)
The operation unit 204 receives user input and outputs the input information to the processing unit 210. For example, the user may operate the operation unit 204 to input information indicating whether to acquire the teacher data or the evaluation data. Further, the user may operate the operation unit 204 to input information indicating a class (correct answer label) corresponding to the time-series data currently acquired by the data acquisition unit 202. The operation unit 204 may be realized by, for example, a mouse, a keyboard, a touch panel, a button, a switch, a lever, a dial, or the like.

(Processing unit 210)
The processing unit 210 acquires teacher data for classification model parameter generation processing and evaluation data for classification processing based on the time-series data acquired by the data acquisition unit 202. In detail, as shown in FIG. 2, the processing unit 210 according to the present embodiment functions as a conversion unit 212, an associating unit 214, a data dividing unit 216, and a separation degree evaluating unit (region specifying unit) 218. Have. The functions of the processing unit 210 will be described below.

-Conversion unit 212-
The transform unit 212 performs Fourier transform, short-time Fourier transform, or wavelet transform on the time-series data (teacher data and evaluation data) from the data acquisition unit 202 to obtain time/frequency/amplitude data (specifically, , A non-negative observation matrix containing non-negative observation vectors (eg, data transformed into a frequency domain representation). For example, as described above, the time/frequency/amplitude data may be a power spectrum, a spectrogram, or a scalogram. Further, the conversion unit 212 may perform Fourier transform or the like on the time-series data of the region (target region) specified by the separation degree evaluation unit 218 described below to generate time/frequency/amplitude data. In the present embodiment, by using the data represented in the frequency domain such as the power spectrum, spectrogram, scalogram, etc., the class to be classified in the processing of the teacher data processing unit 220 or the feature extraction unit 226 described later. The features of can be extracted with higher accuracy.

~Association unit 214~
The associating unit 214 associates the time-series data (teacher data) from the data acquiring unit 202 with the correct answer label (class) input by the user via the operation unit 204. Then, the time series data associated with the correct answer label is converted by the conversion unit 212 described above and output to the teacher data processing unit 220 described below.

-Data division unit 216-
The data division unit 216 divides the time/frequency/amplitude data relating to the teacher data or the evaluation data generated by the conversion in the conversion unit 212 in the frequency domain or the time domain. The data dividing unit 216 can divide the data according to a division rule determined in advance by the user (how many divisions, how large the area should be divided, and the like). By changing to, it becomes possible to extract the features of the class to be classified with higher accuracy or classify with higher accuracy.

~ Separation degree evaluation unit 218 ~
The degree-of-separation evaluation unit 218 evaluates the degree-of-separation indicating the class classification ability in each region divided by the data division unit 216. As the index indicating the degree of separation, for example, the within-class/inter-class variance ratio, the Mahalanobis distance, or the class correct answer rate can be used. Specifically, the degree-of-separation evaluation unit 218 evaluates the degree-of-separation (first evaluation index) of the target class (first class) for each of the regions divided by the data division unit 216 to determine the degree of separation. Area (first area) is extracted, the degree of separation (second evaluation index) of a class (second class) other than the target is evaluated, and the area with high degree of separation (second area) is determined. The target area (predetermined area) from which the teacher data and the evaluation data are acquired is specified based on the extracted areas. More specifically, the degree-of-separation evaluation unit 218 acquires teacher data and evaluation data by, for example, removing an area having a high degree of separation of a non-target class from an area having a high degree of separation of a target class. Identify the target area. That is, in the present embodiment, the classification parameter is not generated using the entire teacher data, and the classification is not performed using the entire evaluation data. In the present embodiment, the classification parameter is generated using the teacher data of the target area having a high target class separation and the non-target class separation not high, and further the classification parameter is generated using the evaluation data of the target area. By performing the above, it is possible to more accurately extract the feature of the target class to be classified or classify the target class with higher precision without being disturbed by the feature of the non-target class.

Furthermore, the teacher data of the target area and the time/frequency/amplitude data of the evaluation data specified by the separation degree evaluation unit 218 are output to the teacher data processing unit 220 and the feature extraction unit 226, respectively, which will be described later.

In addition to the above-described processing, the processing unit 210 needs to perform processing such as processing for subtracting an offset value from the time series data before conversion by the conversion unit 212 and frequency filtering processing for the converted time/frequency/amplitude data. You may go accordingly. The details of the operation of the processing unit 210 will be described later.

(Teacher data processing unit 220)
The teacher data processing unit 220 converts the time/frequency/amplitude data (non-negative first teacher data) (specifically, the teacher observation matrix) of the target area specified by the separation degree evaluation unit 218 received from the processing unit 210. On the other hand, a non-negative matrix factorization is performed to generate a teacher basis matrix and a teacher coefficient matrix (details will be described later). Then, the teacher data processing unit 220 associates the generated teacher base matrix with the correct answer label (class) and outputs the teacher basis matrix to the parameter generation unit 222 and the storage unit 224 described later. The details of the operation of the teacher data processing unit 220 will be described later.

(Parameter generation unit 222)
The parameter generation unit 222 generates a classification model parameter for classifying the time series data into a target class based on the teacher coefficient matrix and the correct answer label described above. The classification model parameter generated by the parameter generation unit 222 may be a parameter according to a classification model in the classification unit 228 described later. For example, the classification model parameter when the classification unit 228 makes a threshold determination may be a threshold. Further, the classification model parameter when the classifying unit 228 makes a linear judgment may be a coefficient of a linear discriminant function that determines a class boundary, and the classification model parameter when the classifying unit 228 makes a quadratic judgment is a class model parameter. It may be a coefficient of a quadratic discriminant function that determines the boundary. The parameter generation unit 222 outputs the classification model parameter generated as described above to the storage unit 224 and stores it.

(Storage unit 224)
The storage unit 224 stores programs and parameters for causing each functional unit of the information processing device 200 to function. For example, the storage unit 224 stores a teacher base matrix generated by the teacher data processing unit 220 performing non-negative matrix factorization on the time/frequency/amplitude data (non-negative first teacher data) of the target region. , And the classification model parameters generated by the parameter generation unit 222.

(Feature extraction unit 226)
The feature extraction unit 226 generates an evaluation coefficient matrix from the time/frequency/amplitude data (non-negative evaluation data) of the target region output by the processing unit 210, based on the teacher basis matrix stored in the storage unit 224, The generated evaluation coefficient matrix is output to the classification unit 228 described later. The details of the operation of the feature extraction unit 226 will be described later.

(Classification unit 228)
The classification unit 228 classifies the evaluation coefficient matrix into a target class based on the classification model parameters stored in the storage unit 224. The classification unit 228 can perform classification based on a classification model parameter according to the classification model using, for example, a coefficient vector included in the evaluation coefficient matrix as a feature vector and using various classification models. Specifically, the classification unit 228 can perform classification using, for example, a classification model based on threshold value determination, linear determination, quadratic determination, SVM (Support Vector Machine), or the like.

(Output unit 230)
The output unit 230 outputs the classification result of the classification unit 228. The output unit 230 can be, for example, a display device such as a display that displays the classification result, or a speaker that outputs the classification result by voice.

<Overview>
The detailed configuration of the information processing apparatus 200 according to the embodiment of the present invention has been described above. Subsequently, before describing an operation example according to the present embodiment, an outline of the present embodiment will be described with reference to FIGS. 3 and 4. FIG. 3 is an explanatory diagram schematically showing a processing example of the NMF according to the present embodiment, and FIG. 4 is a schematic diagram for explaining dimension reduction.

(Nonnegative matrix factorization)
In the present embodiment, a non-negative observation matrix including a non-negative observation vector (data converted into a frequency domain representation) can be obtained by the conversion of the conversion unit 212 described above. The observation matrix is a matrix in which one or more observation vectors obtained from one target section are arranged. Further, in the following description, as described above, the observation matrix acquired as the teacher data may be referred to as a teacher observation matrix, and the observation matrix acquired as the evaluation data may be referred to as an evaluation observation matrix to be distinguished.

Non-negative matrix factorization (hereinafter referred to as NMF) is an algorithm that decomposes one non-negative matrix Y into two non-negative matrices W and H (for example, WH). By decomposing the one non-negative matrix Y into two non-negative matrices W and H, the latent element (feature) of the original non-negative matrix Y can be clarified. Since it is difficult to analytically determine the two non-negative matrices W and H, a method for iteratively obtaining an approximate solution so that an error between Y and WH is a local optimum solution by providing an initial value is proposed. (For example, Non-Patent Documents 1 and 2 above). Note that the local optimum solution generally changes depending on the initial value.

As shown in FIG. 3, for example, a non-negative observation matrix Y (teacher observation matrix or evaluation observation matrix) is decomposed into a product of a non-negative coefficient matrix W and a non-negative basis matrix H by NMF.

Further, as shown in FIG. 3, the observation matrix Y according to the present embodiment is composed of, for example, the m-dimensional observation vector y as a row vector, and the rows are the number n of the target sections in the time-series data. It is a matrix with n rows and m columns.

The basis matrix H according to the present embodiment is obtained by applying NMF to the observation matrix Y, and the decomposed m-dimensional basis vector h is a row vector, and k is composed of as many rows as the number k of basis vectors. It is a matrix with rows and m columns.

The coefficient matrix W according to the present embodiment is, for example, a matrix of n rows and k columns that is configured by k rows of k-dimensional coefficient vectors w and has rows of the number n of observation vectors. Here, the coefficient vector w is a row vector indicating a weighted value indicating how many components of each basis vector included in the basis matrix H are included in a certain observation vector.

As shown in FIG. 3, in the above matrix, there is a relationship of the following mathematical expression (1).

As described above, the basis matrix and the coefficient matrix obtained by decomposing the teacher observation matrix Y _L are the teacher basis matrix H _L and the teacher coefficient matrix W _L , respectively.

In the present embodiment, the teacher data processing unit 220 described above applies NMF to each class of correct labels associated with the teacher observation matrix Y _L to obtain a teacher base matrix H _L and a teacher coefficient matrix W _L. be able to.

For example, the teacher data processing unit 220 first sets a data set (observation vector or observation matrix) Y _C1 , Y _C2 ,..., Y of each class (c1, c2,..., CN) of the teacher observation matrix Y _L. Apply NMF per _CN . Teacher data processor 220, a data set _{Y C1, Y C2, ···,} Y vector or matrix _H C1 which each obtained is decomposed in _CN, H C2, · · _·, a _{H CN,} teachers basis matrix H _L ={H _C1 , H _C2 ,..., H _CN } is generated.

Then, assuming that the coefficient matrix generated from the teacher observation matrix Y _L based on the teacher basis matrix H _L is the teacher coefficient matrix W _L , the teacher observation matrix Y _L , the teacher coefficient matrix W _L , and The relationship of the teacher basis matrix H _L can be expressed by the following mathematical expression (2).

From Equation (2) above, in order to obtain the teacher coefficient matrix W _L , it is necessary to use the inverse matrix of the teacher base matrix H _L. However, since the teacher basis matrix H _L is not generally a regular matrix, it may not have an inverse matrix. Therefore, the teacher data processing unit 220 may generate the teacher coefficient matrix W _L based on, for example, a pseudo inverse matrix of the teacher base matrix (Moore-Penrose pseudo inverse matrix). The pseudo inverse matrix H _L ⁺ of the teacher basis matrix H _L is a k×m matrix, and generally k<m, so the pseudo inverse matrix H _L ⁺ can be expressed by the following mathematical expression (3).

The teacher coefficient matrix W _L can be expressed by the following mathematical expression (4) using the pseudo inverse matrix H _L ⁺ obtained by the mathematical expression (3).

NMF generally has a large amplitude, and it is easy to generate a rule having a characteristic corresponding to data that frequently occurs. Therefore, when NMF is applied to all data sets at once, if there is data with small amplitude or data with low number of appearances depending on the class, it may be difficult to generate a base having features corresponding to the data of the class. is there. Therefore, in the present embodiment, by applying the NMF for each class, a base having a feature corresponding to the data with a small amplitude or the data with a low appearance frequency is easily generated, and the classification accuracy can be further improved. it can. In this embodiment, NMF may be collectively applied to a plurality of classes or data sets of all classes.

The teacher data processing unit 220 described above associates the generated teacher basis matrix H _L with the correct answer label (class) associated by the association unit 214, and causes the parameter generation unit 222 and the storage unit 224 to associate the same. Output.

Next, in the present embodiment, the above-described feature extraction unit 226 generates the evaluation coefficient matrix W _T from the evaluation observation matrix Y _T of the evaluation data, based on the teacher basis matrix H _L stored in the storage unit 224. .. As described above, the basis matrix and the coefficient matrix obtained by decomposing the evaluation observation matrix Y _T are the evaluation basis matrix H _T and the teacher coefficient matrix W _T , respectively. Hereinafter, an example of generating the evaluation coefficient matrix W _T by the feature extraction unit 226 will be described.

The relationship between the teacher basis matrix H _L , the evaluation observation matrix Y _T, and the evaluation coefficient matrix W _T can be expressed by the following Expression (5) based on the above Expression (1).

The evaluation coefficient matrix W _T can be expressed by the following formula (6) by using the pseudo inverse matrix H _L ⁺ obtained by the same method as the formula (3).

Then, in the present embodiment, the feature extraction unit 226 outputs the evaluation coefficient matrix W _T generated as described above to the classification unit 228, and the classification unit 228 stores the evaluation coefficient matrix W _T in the storage unit 224. Classify based on the stored classification model parameters.

In the present embodiment, as described above, the classification is performed by using the coefficient vector indicating the weight with respect to the basis matrix as the feature vector, rather than the direct classification based on the magnitude of the power of the specific frequency, for example. The accuracy can be improved.

Further, in the present embodiment, the classification process is performed by using the evaluation coefficient matrix W _T generated from the evaluation data, as compared with the case where the classification process is performed using the evaluation data (time series data) itself. The dimension of the feature vector in the processing is reduced, and the processing amount of the classification processing can be suppressed.

For example, as shown in FIG. 4, by NMF, the evaluation data G20 which is a power spectrum can be represented by the sum of the product of the basis vector G21 and the coefficient W _{1 and} the product of the basis vector G22 and the coefficient W _2. .. In the example of FIG. 4, when performing the classification process using the evaluation data G20 as a feature vector, it is necessary to perform processing in a dimension according to the frequency resolution of the evaluation data G20. On the other hand, in the example of FIG. 4, the dimension of the coefficient vector is 2. Therefore, as described above, since the dimension number of the coefficient vector is smaller than the dimension of the evaluation data G20, the dimension of the feature vector in the classification processing is reduced by performing the classification processing using the evaluation coefficient matrix W _T. The processing amount of the classification processing can be suppressed.

(Target area)
However, although the amount of classification processing can be suppressed by using NMF, since the spectrogram and scalogram are data having two axes of time and frequency, when NMF is applied to a spectrogram, etc. There may be a limit to the suppression of.

In addition, as described above, in general, the NMF is likely to select a value of amplitude or a base with a high appearance frequency. Therefore, when NMF is applied to the entire time/frequency/amplitude data, for example, a base with a low class separation ability and a large amplitude is selected, and a base with a high class separation capacity and a small amplitude is selected depending on the data. It may be lost. In such a case, since the separation accuracy may decrease, there is a limit in improving the classification accuracy.

It will be described with reference to FIG. 5 that there may be a limit of classification accuracy. FIG. 5 is a schematic diagram showing an example in which NMF is applied to the entire time/frequency/amplitude data. The spectrograms G61 and G62 shown in FIG. 5 are spectrograms showing the characteristics of class 1 and class 2, respectively. Furthermore, in the spectrograms G61 and G62, the darkness of black indicates the magnitude of the amplitude.

In the example shown in FIG. 5, the magnitude of the amplitude in the region R1 where the amplitude is large appears commonly in the spectrogram G61 showing the features of class 1 and the spectrogram G62 showing the features of class 2, and is not a feature effective for classification. Make it not exist. On the other hand, in the example shown in FIG. 15, it is assumed that the magnitudes of the amplitudes in the regions R2 and R3 where the amplitude is small are effective characteristics for class 1 and class 2.

However, if NMF is applied to the entire data of the spectrograms G61 and G62, the bases G611 and G621 corresponding to the region R1 having a large amplitude are selected, and the amplitudes that are effective for the classification of class 1 and class 2 are selected. Small bases of may not be selected. In such a case, the coefficient G613 obtained from the spectrogram G61 and the coefficient G623 obtained from the spectrogram G62 are the same, which makes it difficult to accurately classify the class 1 and the class 2.

Therefore, in the present embodiment, in order to further improve the accuracy of classification, NMF is partially applied only in the time domain and frequency domain effective for class classification. Such an application example will be described with reference to FIG. FIG. 6 is a schematic diagram showing an example in which NMF is applied to the time/frequency/amplitude data portion.

The spectrograms G71 and G72 shown in FIG. 6 are spectrograms showing the characteristics of class 1 and class 2, respectively, similarly to the spectrograms G61 and G62 shown in FIG. As shown in FIG. 6, when the NMF is partially applied only to the region R4 of the spectrograms G71 and G72, the bases G711 and G721 corresponding to the region R2 and the bases G712 and G722 corresponding to the region R3 are selected. .. As a result, the coefficient G713 and the coefficient G714 obtained from the spectrogram G71 are greatly different from the coefficient G723 and the coefficient 724 obtained from the spectrogram G72, and the class 1 and the class 2 can be easily classified. That is, the accuracy of classification can be further improved by partially applying NMF only to the time domain and frequency domain effective for class classification. Further, in the present embodiment, since NMF may be applied only to the region R4, it is possible to further reduce the processing amount.

For example, tapping data is generally data in a limited time zone, and it is sufficient to perform analysis according to the purpose by limiting it to a part of the time zone or frequency zone where a predetermined amplitude appears in the time/frequency/amplitude data. May be. Further, even for data such as vibration noise of a machine that outputs a sound in a specific frequency band, it may be sufficient to perform analysis according to the purpose by limiting the frequency band.

Further, in the present embodiment, in order to further improve the accuracy of classification, not only the time domain and frequency domain effective for the target class classification but also the time domain and frequency effective for the non-target class classification are used. The NMF is partially applied after excluding the region. That is, in the present embodiment, by applying the NMF only to the region where the target class separation is high and the non-target class separation is not high, the characteristics of the non-target class are disturbed. The feature of the desired class to be classified can be extracted with higher accuracy, or can be classified into the target class with higher accuracy. As a result, in this embodiment, the accuracy of classification can be further improved. In addition, in the present embodiment, since it is sufficient to limit the area and apply the NMF, it is possible to further suppress the processing amount.

In the following description, the target class classification is classified into, for example, two classes (classes related to the state of the measurement object (machine)) indicating whether the measurement object is normal or abnormal. That is. Further, as the class classification other than the purpose, for example, a class relating to individual difference of the measurement object (machine) or a class relating to environment difference in which the measurement object (machine) is installed is classified. That is, in the present embodiment, by partially applying the NMF as described above, for example, the influence of the individual difference or the environmental difference of the measurement target is further suppressed, and the target measurement target is normal. It is possible to highly accurately extract the characteristics of the class indicating whether the class is present or abnormal.

(Separation degree)
Then, in the present embodiment, the region (target region) to which NMF is applied is limited based on the degree of separation of the target class classification and the degree of separation of the non-target class classification. The limitation of the target area to which the NMF according to the present embodiment is applied will be described with reference to FIGS. 7 to 9. FIG. 7 is an explanatory diagram showing an example of data division according to the present embodiment, and more specifically, an example of data division by the data division unit 216 and acquisition of teacher data and evaluation data by the separation degree evaluation unit 218. It is an explanatory view shown. 8 and 9 are explanatory diagrams showing an example of acquisition of teacher data and evaluation data according to the present embodiment.

As shown in FIG. 7, the data dividing unit 216, for example, the area A0 (in FIGS. 8 and 9) of the time/frequency/amplitude data (observation matrix of the non-negative second teacher data) obtained from the teacher data. (Corresponding to the region R80), the data is divided into four regions A1, A2, A3 and A4. The number of divisions by the data division unit 216 may be an arbitrary number of 2 or more.

Then, the degree-of-separation evaluation unit 218, for example, with respect to each of the divided areas A1, A2, A3, and A4, the degree-of-separation (first evaluation index) between classes of the target class (first class). As a result, the intra-class/inter-class variance ratio is calculated as the degree of separation. First, the within-class variance σ _w ² and the between-class variance σ _B ² can be expressed by the following formulas (7) and (8).

However, the equation (7), in (8), n is the total number of data, n _i is the number of data of the class i, x is a feature vector, m _i is the average of the feature vectors of the class i, m is the total feature value It is the average of the vectors. Using the within-class variance σ _w ² and the between-class variance σ _B ² , the within-class/inter-class variance ratio J is expressed by the following mathematical expression (9).

It can be considered that the larger the above-mentioned in-class/inter-class dispersion ratio J, the greater the degree of separation and the better the feature amount. The index indicating the degree of separation is not limited to the above, and for example, the Mahalanobis distance or the like can be used.

The degree-of-separation evaluation unit 218 calculates the degree of separation for the target class as described above in each of the four areas A1, A2, A3, and A4. Further, the separation degree evaluation unit 218, for example, an area (first area) in which the value of the separation degree calculated from the four areas A1, A2, A3, and A4 is equal to or greater than a preset threshold L ₁ (first threshold). ) Extract R10 (see FIG. 8).

In the same manner as described above, the separability evaluation unit 218, for example, each area A1, A2, A3 of the time/frequency/amplitude data (observation matrix of the non-negative third teacher data) obtained from the divided teacher data, For A4, the degree of separation (second evaluation index) for the non-target class (second class) is calculated. Further, the separation degree evaluation unit 218, for example, an area (second area) in which the value of the separation degree calculated from the four areas A1, A2, A3, and A4 is equal to or greater than a preset threshold L ₂ (second threshold). ) Extract R12 (see FIG. 8). It should be noted that the number of non-purpose classes may be two or more, and in this case, the degree-of-separation evaluation unit 218 may determine, for example, for each of the divided regions A1, A2, A3, and A4 about other non-purpose classes. Is calculated, and the region R14 (see FIG. 9) in which the value of the separation calculated from the four regions A1, A2, A3, and A4 is equal to or greater than the preset threshold L ₃ is extracted.

Further, the separation degree evaluation unit 218 identifies a target area (predetermined area) to which NMF is applied based on the extracted areas R10, R12, and R14. For example, the degree-of-separation evaluation unit 218 identifies the target region to which NMF is applied, based on any of the sum, difference, or product of the extracted regions R10, R12, and R14. Specifically, as illustrated in FIG. 8, the separation degree evaluation unit 218 identifies a region G810 to which NMF is applied by removing the region R12 having a threshold value L ₂ or more from the region R10 having a threshold value L ₁ or more. .. When there are two non-objective classes, as shown in FIG. 9, the separation degree evaluation unit 218 determines whether the separation degree evaluation unit 218 starts from the region R10 having a threshold value L ₁ or more to the region R12 having a threshold value L ₂ or more and a threshold value L ₃ or more. A region G820 to which NMF is applied is specified by removing the region R14 of. That is, in the present embodiment, by selecting a region in which the degree of separation of the target class is high and the degree of separation of the non-target class is not high, the target region effective in classifying the target class is selected. Can be found efficiently.

Then, in the present embodiment, as described above, by applying NMF to the target region obtained by removing the region having a high degree of separation of the non-target class from the region having a high degree of separation of the target class, a higher separation is achieved. It is possible to obtain a feature quantity that has a degree and is effective for more accurate classification.

Then, in the learning phase, the degree-of-separation evaluation unit 218 sets the data in the time domain/frequency domain in the target area to which the NMF specified as described above is applied to the teacher data (the teacher of the non-negative first teacher data). Observation matrix). Further, the degree-of-separation evaluation unit 218 specifies the time domain/frequency domain identified as described above as a data domain (target domain) for obtaining the evaluation data, and the corresponding time/frequency/amplitude of the data domain. Data is acquired as evaluation data (evaluation observation matrix of non-negative evaluation data).

Further, in the present embodiment, the class classification correct answer rate may be used as the degree of separation. In detail, the separability evaluation unit 218 applies NMF to each of the divided data as shown in FIG. 7, and acquires a base matrix and a coefficient matrix in a plurality of ranks. Note that the time/frequency/amplitude data may be converted into a one-dimensional column vector in order to apply NMF.

In the following example, a case where two classes C1 and C2 exist will be described. First, the basis matrices H _{C1, A1} , H _{C1, A2} , H _{C1, A3} , H _{C1, A4} , and the coefficient matrix are calculated from a plurality of time/frequency/amplitude data regions A1, A2, A3, A4 belonging to the class C1. It is possible to obtain W _C1,A1 , W _C1,A2 , W _C1,A3 , W _C1,A4 , respectively. Then, from the plurality of time/frequency/amplitude data regions A1, A2, A3, and A4 belonging to the class C2, the basis matrices H _{C2, A1} , H _{C2, A2} , H _{C2, A3} , H _{C2, A4} , And a coefficient matrix W _C2,A1 , W _C2,A2 , W _C2,A3 , W _C2,A4 , respectively. The rank indicating the number of decomposed bases ranges from 1 to the number of dimensions of the original data.

Then, as described above, the classification accuracy rate of the class C1 or the class C2 is used as the index indicating the degree of separation of the class C1 or the class C2. First, the degree-of-separation evaluation unit 218 combines the bases H _C1,A1 , H _C2,A1 obtained from the classes C1 and C2, respectively, to obtain a common base H _C1-C2,A1 ={H _C1,A1 , H _{C2 , A1} } is calculated. Data _{Y C1} of the original class _{C1, A0,} data _{Y C2} class _C2, the _A0 applying _{H C1-C2,} pseudo-inverse _{H C1-C2} of _{A1, A1} ^+, the coefficient matrix _{W C1} of each _{class, A0} , _{WC2, A0} can be expressed by the following mathematical expressions (10) and (11).

The separation degree evaluation unit 218 classifies the coefficient matrices W _C1,A0 , W _C2,A0 as the feature vectors of the class C1 and the class C2, respectively, and calculates the classification accuracy rate. Here, the classification method performed by the separation degree evaluation unit 218 may be the same as the classification method of the classification unit 228, for example.

The data dividing unit 216 may further recursively divide the data in each of the four areas A1, A2, A3, and A4 shown in FIG. For example, the area A1 shown in FIG. 7 may be divided into four areas A11, A12, A13, and A14. Further, the separation degree evaluation unit 218 calculates the separation degree for each of the divided data and the data obtained by combining the divided data. The above recursive division may be repeated until the time domain/frequency domain becomes the minimum unit, or until the size of the divided data reaches a predetermined value.

In the above description, an example of recursive division has been described in order to efficiently extract the time domain/frequency domain having a high degree of separation, but the present invention is not limited to this example. For example, the degree of separation in part or all of the time domain/frequency domain may be evaluated, or the time domain/frequency domain with high degree of separation may be extracted by another method.

<Operation example>
The outline of the present embodiment has been described above. Next, an example of the operation (classification method) of this embodiment will be described with reference to FIGS. 10 and 11, separately for the learning phase and the classification phase. FIG. 10 is a flowchart showing an operation example of this embodiment in the learning phase, and FIG. 11 is a flowchart showing an operation example of this embodiment in the classification phase.

(Example of operation in the learning phase)
Next, details of the operation example of the learning phase according to the present embodiment will be described. As shown in FIG. 10, the learning phase includes, for example, a plurality of steps from step S101 to step S125. The details of each step included in the learning phase will be described below.

~ Step S101 ~
The data acquisition unit 202 acquires time series data from the sensor 100. At this time, the processing unit 210 may perform processing of subtracting an offset value from the acquired time series data. Alternatively, the processing unit 210 may perform frequency filtering processing or the like on the time/frequency/amplitude data converted in step S107 described later. By doing so, the accuracy of the time/frequency/amplitude data that is the target of NMF and the effectiveness of the feature quantity effective for classification can be further enhanced.

~ Step S103 ~
Subsequently, the associating unit 214 associates the time-series data acquired in step S101 with the correct label (target class and non-target class) input by the user via the operation unit 204.

~ Step S105
The processing unit 210 determines whether or not the acquisition of the time series data used in the learning phase is completed. For example, the determination may be performed based on, for example, the input of the operation unit 204. Then, if the acquisition of the time-series data used in the learning phase has not been completed, the process returns to step S101. On the other hand, when the acquisition of the time series data used in the learning phase has been completed, the process proceeds to step S107.

~ Step S107 ~
The conversion unit 212 converts the time series data acquired in step S103 into time/frequency/amplitude data.

~ Step S109 ~
Subsequently, the data division unit 216 divides the time/frequency/amplitude data acquired in step S107 in the time domain and the frequency domain.

~ Step S111 ~
Next, the degree-of-separation evaluation unit 218 evaluates the degree-of-separation of the target class for each area divided by the data division unit 216 in step S109.

~ Step S113 ~
The degree-of-separation evaluation unit 218 extracts an area having a high degree of separation evaluated in step S111 from the plurality of areas divided in step S109.

~ Step S115 ~
Subsequently, the separation degree evaluation unit 218 evaluates the separation degree of the non-target class for each of the areas divided by the data division unit 216 in step S109.

~Step S117~
The degree-of-separation evaluation unit 218 extracts an area having a high degree of separation evaluated in step S115 from the plurality of areas divided in step S109.

~Step S119~
The degree-of-separation evaluation unit 218 identifies the target area by excluding the area with high degree of separation of the non-target class extracted in step S117 from the area with high degree of separation of the target class extracted in step S113. To do.

~ Step S121 ~
The teacher data processing unit 220 acquires the time/frequency/amplitude data (teacher observation matrix of teacher data) of the target region specified in step S119.

~ Step S123 ~
The teacher data processing unit 220 acquires the teacher base matrix and the teacher coefficient matrix by applying NMF to the time/frequency/amplitude data (teacher observation matrix of teacher data) of the target area acquired in step S121.

~ Step S125
Further, the parameter generation unit 222 generates a classification model parameter based on the teacher coefficient matrix acquired in step S123 and the correct answer label (target class) associated in step S103.

The operation example of this embodiment in the learning phase has been described above. Next, an operation example of the present embodiment in the classification phase using the classification model parameter generated in the learning phase will be described with reference to FIG.

(Operation example of classification phase)
Next, details of an operation example of the classification phase will be described. As shown in FIG. 11, the classification phase according to this embodiment includes, for example, a plurality of steps from step S201 to step S215. The details of each step included in the classification phase will be described below.

~ Step S201 ~
The data acquisition unit 202 acquires time series data. At this time, the processing unit 210 may perform processing of subtracting an offset value from the acquired time series data. Alternatively, the processing unit 210 may perform frequency filtering processing or the like on the time/frequency/amplitude data converted in step S107 described later. By doing so, the accuracy of the time/frequency/amplitude data that is the target of NMF and the accuracy of classification can be further improved.

~Step S203~
The conversion unit 212 converts the time series data acquired in step S201 into time/frequency/amplitude data.

~ Step S205
Subsequently, the data division unit 216 divides the time/frequency/amplitude data acquired in step S203 in the time domain and the frequency domain.

~Step S207~
Subsequently, the separability evaluation unit 218 identifies the same region as the target region identified in step S119 of the learning phase described with reference to FIG. 10 from the time/frequency/amplitude data divided in step S205. ..

~ Step S209 ~
Further, the feature extraction unit 226 acquires time/frequency/amplitude data (evaluation observation matrix of evaluation data) of the target area specified in step S207.

~ Step S211 ~
Subsequently, the feature extraction unit 226 extracts the time/frequency/amplitude data (evaluation observation matrix of the evaluation data) of the target region acquired in step S209 based on the teacher base matrix acquired in step S123 of the learning phase described above. Generate an evaluation coefficient matrix.

~Step S213~
Subsequently, the classification unit 228 classifies the evaluation coefficient matrix generated in step S211 based on the classification model parameters generated in step S125 of the learning phase described above.

~ Step S215 ~
Finally, the output unit 230 outputs the classification result in step S213.

<<Summary>>
As described above, in the embodiment of the present invention, in order to further improve the accuracy of the classification, not only the time domain and the frequency domain effective for the target class classification but also the non-target class classification The NMF is partially applied after excluding the effective time domain and frequency domain. That is, according to the present embodiment, by applying the NMF only to the region where the target class separation is high and the non-target class separation is not high, the characteristics of the non-target class are disturbed. Without extracting the features that are effective in classifying the target class that you want to classify with higher accuracy (in other words, it becomes easier to select the bases that are effective in classifying the target class), and with higher accuracy in the target class. Can be classified into. As a result, in this embodiment, the accuracy of classification can be further improved. In addition, in the present embodiment, since it is sufficient to limit the area and apply the NMF, it is possible to further suppress the processing amount.

Further, in the above-described embodiment, based on the acoustic signal (time-series data) obtained by collecting the hitting sound when hitting the object to be measured (for example, concrete) with a hitting tool such as a hammer. An example in which the measurement target is classified into two classes (first target class) indicating whether the measurement target is normal or abnormal has been described. However, the embodiment of the present invention is not limited to the above-mentioned example, and may be classified into, for example, three or more classes, and time series data of vibration of a measurement object (for example, a machine tool). Based on, the state and ability of the measurement object may be classified into a plurality of classes. Further, in the embodiment of the present invention, the target time-series data is not particularly limited, and may be image data that changes with time, data such as environmental temperature, flow rate, or the like. Data (for example, voltage value, current value, etc.), motion data of measurement target (for example, building, machine, person, ball, moving body (automobile), etc.), various biological information obtained from human body (For example, pulse, heartbeat, body temperature, brain wave, etc.) may be used. That is, the embodiments of the present invention can be applied to various machines, human diagnosis, and the like.

<<Hardware configuration>>
The embodiments of the present invention have been described above. Information processing such as the above-described preprocessing, teacher data processing, parameter generation processing, feature extraction processing, and classification processing is realized by cooperation of software and hardware of the information processing apparatus 200. Hereinafter, a hardware configuration example of the information processing apparatus 1000 will be described as a hardware configuration example of the information processing apparatus 200 that is the information processing apparatus 200 according to the embodiment of the present invention.

FIG. 12 is an explanatory diagram showing a hardware configuration example of the information processing apparatus 1000 according to the embodiment of the present invention. As illustrated in FIG. 12, the information processing device 1000 includes a CPU (Central Processing Unit) 1001, a ROM (Read Only Memory) 1002, a RAM (Random Access Memory) 1003, an input device 1004, and an output device 1005. The storage device 1006 and the communication device 1007 are provided.

The CPU 1001 functions as an arithmetic processing unit and a control unit, and controls overall operations in the information processing apparatus 1000 according to various programs. Further, the CPU 1001 may be a microprocessor. The ROM 1002 stores programs used by the CPU 1001 and calculation parameters. The RAM 1003 temporarily stores a program used in the execution of the CPU 1001 and parameters that appropriately change in the execution. These are connected to each other by a host bus composed of a CPU bus and the like. The functions of, for example, the processing unit 210, the teacher data processing unit 220, the parameter generation unit 222, the feature extraction unit 226, the classification unit 228 and the like are realized mainly by the cooperation of the CPU 1001, the ROM 1002, the RAM 1003, and the software.

The input device 1004 includes an input unit such as a mouse, a keyboard, a touch panel, a button, a microphone, a switch, and a lever for the user to input information, and an input control circuit that generates an input signal based on the input by the user and outputs the input signal to the CPU 1001. Etc. By operating the input device 1004, the user of the information processing apparatus 1000 can input various data to the information processing apparatus 1000 and instruct a processing operation. The input device 1004 corresponds to the operation unit 204.

The output device 1005 includes, for example, a liquid crystal display (LCD) device, an OLED (Organic Light Emitting Diode) device, and a display device such as a lamp. Further, the output device 1005 includes an audio output device such as a speaker and headphones. For example, the display device displays a captured image, a generated image, or the like. On the other hand, the voice output device converts voice data and the like into voice and outputs the voice. The output device 1005 corresponds to the output unit 230.

The storage device 1006 is a device for storing data. The storage device 1006 may include a storage medium, a recording device that records data in the storage medium, a reading device that reads data from the storage medium, a deletion device that deletes data recorded in the storage medium, and the like. The storage device 1006 stores programs executed by the CPU 1001 and various data. The storage device 1006 corresponds to the storage unit 224.

The communication device 1007 is, for example, a communication interface including a communication device for connecting to a communication network. The communication device 1007 may include a wireless LAN (Local Area Network) compatible communication device, an LTE (Long Term Evolution) compatible communication device, a wire communication device that performs wired communication, or a Bluetooth (registered trademark) communication device.

<<Supplement>>
The preferred embodiments of the present invention have been described above in detail with reference to the accompanying drawings, but the present invention is not limited to these examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

For example, in the above embodiment, an example in which the same device (information processing device 200) acquires the teacher data and the evaluation data has been described, but the present invention is not limited to this example. For example, the device that acquires the teacher data and the evaluation data may be separate devices, and the teacher data and the correct answer label may be stored in the storage unit 224 in advance. That is, in the above-described embodiment, an example has been described in which the information processing device 200 has a function as a parameter generation device that generates classification model parameters and a function as a classification device that classifies time-series data into a plurality of classes. However, the present invention is not limited to such an example. For example, in the present embodiment, the parameter generation device that performs the classification model parameter generation process and the classification device that performs the classification process using the classification model parameter generated by the parameter generation device may be provided as different devices. Good.

Further, the above-described embodiment is, for example, a program for causing a computer to function as the information processing apparatus 200 according to the present embodiment (for performing the classification method according to the present embodiment), and a non-transitory program in which the program is recorded. It may include a tangible medium. Further, the program may be distributed via a communication line (including wireless communication) such as the Internet.

Further, the processing steps in the operation (classification method) of the present embodiment do not necessarily have to be executed in time series in the order described in the flowchart. For example, in the present embodiment, the processing steps may be executed in an order different from the order described as the flowchart, or may be executed in parallel.

10 information processing system 100 sensor 200, 1000 information processing device 200a cloud 200b IoT gateway 200c edge terminal 202 data acquisition unit 204 operation unit 210 processing unit 212 conversion unit 214 associating unit 216 data division unit 218 separation degree evaluation unit 220 teacher data processing Part 222 Parameter generation part 224 Storage part 226 Feature extraction part 228 Classification part 230 Output part 1001 CPU
1002 ROM
1003 RAM
1004 input device 1005 output device 1006 storage device 1007 communication device

Claims (4)

A classifying device for classifying time series data into a desired first class,
From the time series data expressed in at least the frequency domain or the time domain, a processing unit that acquires non-negative evaluation data in a predetermined data domain,
A teacher coefficient matrix generated by performing non-negative matrix factorization on the non-negative first teacher data corresponding to the predetermined data area, and the first teacher data associated with the first teacher data. A storage unit for storing classification model parameters generated based on the class of
A feature extraction unit that generates an evaluation coefficient matrix from the evaluation data based on a teacher basis matrix generated by performing non-negative matrix factorization on the first teacher data;
A classification unit that classifies the calculated evaluation coefficient matrix into the first class based on the classification model parameters;
An area specifying unit for specifying the predetermined data area,
Equipped with
The area specifying unit,
Each non-negative second teacher data associated with each of the first classes is divided in the frequency domain or the time domain,
Calculating a first evaluation index indicating the degree of separation between the first classes for each of the divided regions, based on the observation matrix of each of the second teacher data for each of the divided regions,
Each non-negative third teacher data associated with each second target class is divided in the frequency domain or the time domain,
Calculating a second evaluation index indicating the degree of separation between the second classes for each of the divided regions, based on the observation matrix of each of the third teacher data for each of the divided regions,
The first evaluation index is higher than a preset first threshold value, the first area is specified by extracting the divided area,
The second area is specified by extracting the divided area in which the second evaluation index is higher than a preset second threshold value,
Specifying the predetermined data area based on the specified first and second areas;
Classifier. A classifying device for classifying the time series data into a target first class;
A sensor for acquiring the time series data,
A classification system including
The classification device is
From the time series data expressed in at least the frequency domain or the time domain, a processing unit that acquires non-negative evaluation data in a predetermined data domain,
A teacher coefficient matrix generated by performing non-negative matrix factorization on the non-negative first teacher data corresponding to the predetermined data area, and the first teacher data associated with the first teacher data. A storage unit for storing classification model parameters generated based on the class of
A feature extraction unit that generates an evaluation coefficient matrix from the evaluation data based on a teacher basis matrix generated by performing non-negative matrix factorization on the first teacher data;
A classification unit that classifies the calculated evaluation coefficient matrix into the first class based on the classification model parameters;
An area specifying unit for specifying the predetermined data area,
Have
The area specifying unit,
Each non-negative second teacher data associated with each of the first classes is divided in the frequency domain or the time domain,
Calculating a first evaluation index indicating the degree of separation between the first classes for each of the divided regions, based on the observation matrix of each of the second teacher data for each of the divided regions,
Each non-negative third teacher data associated with each second target class is divided in the frequency domain or the time domain,
Calculating a second evaluation index indicating the degree of separation between the second classes for each of the divided regions, based on the observation matrix of each of the third teacher data for each of the divided regions,
The first evaluation index is higher than a preset first threshold value, the first area is specified by extracting the divided area,
The second area is specified by extracting the divided area in which the second evaluation index is higher than a preset second threshold value,
Specifying the predetermined data area based on the specified first and second areas;
Classification system. A classification method for classifying time series data into a desired first class,
From the time series data expressed in at least the frequency domain or the time domain, obtaining non-negative evaluation data in a predetermined data domain,
A teacher coefficient matrix generated by performing non-negative matrix factorization on the non-negative first teacher data corresponding to the predetermined data area, and the first teacher data associated with the first teacher data. Storing classification model parameters generated based on the class of
Generating an evaluation coefficient matrix from the evaluation data based on a teacher basis matrix generated by performing non-negative matrix factorization on the first teacher data;
Classifying the calculated evaluation coefficient matrix into the first class based on the classification model parameters;
Specifying the predetermined data area,
Including
Specifying the predetermined data area is
Each non-negative second teacher data associated with each of the first classes is divided in the frequency domain or the time domain,
Calculating a first evaluation index indicating the degree of separation between the first classes for each of the divided regions, based on the observation matrix of each of the second teacher data for each of the divided regions,
Each non-negative third teacher data associated with each second target class is divided in the frequency domain or the time domain,
Calculating a second evaluation index indicating the degree of separation between the second classes for each of the divided regions, based on the observation matrix of each of the third teacher data for each of the divided regions,
The first evaluation index is higher than a preset first threshold value, the first area is specified by extracting the divided area,
The second area is specified by extracting the divided area in which the second evaluation index is higher than a preset second threshold value,
Specifying the predetermined data area based on the specified first and second areas;
Having
Classification method. A program for classifying time series data into a desired first class,
On the computer,
From the time series data expressed in at least the frequency domain or the time domain, a function of acquiring non-negative evaluation data in a predetermined data domain,
A teacher coefficient matrix generated by performing non-negative matrix factorization on the non-negative first teacher data corresponding to the predetermined data area, and the first teacher data associated with the first teacher data. The ability to store the classification model parameters generated based on the class of
A function of generating an evaluation coefficient matrix from the evaluation data based on a teacher basis matrix generated by performing non-negative matrix factorization on the first teacher data;
A function of classifying the calculated evaluation coefficient matrix into the first class based on the classification model parameter;
A function for specifying the predetermined data area,
Is realized,
In the function of specifying the predetermined data area,
Each non-negative second teacher data associated with each of the first classes is divided in the frequency domain or the time domain,
Calculating a first evaluation index indicating the degree of separation between the first classes for each of the divided regions, based on the observation matrix of each of the second teacher data for each of the divided regions,
Each non-negative third teacher data associated with each second target class is divided in the frequency domain or the time domain,
Calculating a second evaluation index indicating the degree of separation between the second classes for each of the divided regions, based on the observation matrix of each of the third teacher data for each of the divided regions,
The first evaluation index is higher than a preset first threshold value, the first area is specified by extracting the divided area,
The second area is specified by extracting the divided area in which the second evaluation index is higher than a preset second threshold value,
Specifying the predetermined data area based on the specified first and second areas;
Do things,
program.

Images