JP2021135357A - Classification system, information processing device, classification method and program - Google Patents

Abstract

To provide a classification system capable of appropriately performing feature analysis processing even in the case of suppressing the capacity of a storage area, an information processing device, a classification method and a program.SOLUTION: A classification system performs feature extraction and class classification based on a time frequency analysis. The classification system includes: a learning part including time frequency analysis processing means for converting a time waveform showing time-series amplitude of a known signal having a known signal source into data showing time-series frequency and signal strength, feature analysis processing means for using classification matrix operation to the data to generate a feature amount and a feature extraction parameter, and learning processing means for estimating a classification parameter from the feature amount; and a classification part including feature extraction processing means for calculating a feature amount of an unknown signal having an unknown signal source with reference to the feature extraction parameter, and classification processing means for estimating a signal source of the unknown signal from the feature amount of the unknown signal with reference to the classification parameter.SELECTED DRAWING: Figure 1

Description

The present invention relates to a classification system, an information processing apparatus, a classification method for performing feature extraction based on time-frequency analysis and classification of signal sources, and a program for causing a computer to execute the classification method.

Conventionally, various methods have been proposed for classifying acoustic signals output from a plurality of signal sources according to the signal sources (see, for example, Non-Patent Document 1). For example, a classification system that classifies acoustic signals from a plurality of signal sources performs time-frequency analysis processing on the time waveform of the acoustic signal. Then, the classification system performs feature analysis processing on the matrix data obtained as the analysis result by using matrix calculation, and acquires the feature amount included in the time waveform.

Keigo Kimura, Yuzuru Tanaka, and Mineichi Kudo, “A Fast Hierarchical Alternating Least Squares Algorithm for Orthogonal Nonnegative Matrix Factorization”, JMLR 39 (2014), pp.129-141

However, the conventional classification system requires a storage area that can store both the matrix data used when performing the matrix operation in the feature analysis process and the matrix data that is the operation result obtained by the matrix operation. .. In particular, in order to improve the classification ability, it is necessary to handle matrix data having a large data size, and a storage area having a correspondingly large capacity is required. When the capacity of the storage area is small, the matrix data required for the matrix operation cannot be stored in the storage area, and the feature analysis process is interrupted. Therefore, it is difficult to properly perform the feature analysis process. ..

Therefore, there is a demand for a classification system, an information processing device, a classification method, and a program that can appropriately perform feature analysis processing even when the capacity of the storage area is suppressed.

The classification system according to the present invention is a classification system that performs feature extraction and classification based on time-frequency analysis, and obtains a time waveform indicating the time-series amplitude of a known signal whose signal source is known, and a time-series frequency. And the time-frequency analysis processing means that converts the data into data indicating the signal strength, the feature analysis processing means that generates the feature amount and the feature extraction parameter by using the partial matrix calculation on the data, and the classification parameter is estimated from the feature amount. With reference to the learning unit having the learning processing means to be used, the feature extraction processing means for calculating the feature amount of the unknown signal whose signal source is unknown by referring to the feature extraction parameter, and the classification parameter with reference to the unknown signal. It is provided with a classification unit having a classification processing means for estimating a signal source of the unknown signal from a feature amount.

The information processing device according to the present invention is an information processing device that performs feature extraction and classification based on time-frequency analysis, and obtains a time-series time waveform indicating the time-series amplitude of a known signal whose signal source is known. Time-frequency analysis processing means that converts the data into data indicating the frequency and signal strength of The learning processing means for estimating the above, the feature extraction processing means for calculating the feature amount of the unknown signal whose signal source is unknown by referring to the feature extraction parameter, and the classification parameter with reference to the feature amount of the unknown signal. It is provided with a classification processing means for estimating the signal source of the unknown signal.

The classification method according to the present invention is a classification method in which feature extraction and classification are performed based on time-frequency analysis, and a time waveform indicating the time-series amplitude of a known signal whose signal source is known is a time-series frequency. And the data indicating the signal strength, the feature quantity and the feature extraction parameter are generated for the data by using the submatrix calculation, the classification parameter is estimated from the feature quantity, the feature extraction parameter is referred to, and the signal is used. The feature amount of the unknown signal whose source is unknown is calculated, the classification parameter is referred to, and the signal source of the unknown signal is estimated from the feature amount of the unknown signal.

The program according to the present invention uses a computer that performs feature extraction and classification based on time-frequency analysis, a time waveform showing the time-series amplitude of a known signal whose signal source is known, and a time-series frequency and signal strength. A time-frequency analysis processing means for converting into the data to be shown, a feature analysis processing means for generating a feature amount and a feature extraction parameter from the data by using a partial matrix calculation, a learning processing means for estimating a classification parameter from the feature amount, the above. A feature extraction processing means that calculates the feature amount of an unknown signal whose signal source is unknown by referring to the feature extraction parameter, and a classification that estimates the signal source of the unknown signal from the feature amount of the unknown signal by referring to the classification parameter. It is intended to function as a processing means.

According to the present invention, by using the submatrix operation in the feature analysis process, the storage area used in the operation can be made smaller than that in the case of using the matrix operation, so that the capacity of the storage area is suppressed. Even in this case, the feature analysis process can be appropriately performed.

It is a block diagram which shows an example of the structure of the classification system which concerns on Embodiment 1. FIG. It is a hardware block diagram which shows an example of the structure of the learning part of FIG. It is a hardware block diagram which shows another example of the structure of the learning part of FIG. It is a block diagram which shows an example of the structure of the conventional classification system. It is a schematic diagram which shows an example of the result of the time frequency analysis processing by the time frequency analysis processing means. It is a schematic diagram which shows an example of the result of the feature analysis processing by the feature analysis processing means. It is a schematic diagram which shows an example of the result of the learning processing by a learning processing means. It is the schematic for demonstrating the sum operation processing of a matrix. It is the schematic for demonstrating the product operation process of a matrix. It is the schematic for demonstrating the 1st problem caused by lack of a main storage area. It is the schematic for demonstrating the 2nd problem caused by lack of a main storage area. It is the schematic for demonstrating a submatrix. It is a schematic diagram for demonstrating sum arithmetic processing using a submatrix. It is a block diagram which shows an example of the structure of the submatrix sum calculation means of FIG. It is a schematic diagram which shows an example of the state of the main storage area and the auxiliary storage area of the sum operation processing using a submatrix. It is a schematic diagram for demonstrating the product operation processing using a submatrix. It is a block diagram which shows an example of the structure of the partial matrix product calculation means of FIG. It is a schematic diagram which shows an example of the state of the main storage area and the auxiliary storage area of the product operation processing using a submatrix.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. The present invention is not limited to the following embodiments, and can be variously modified without departing from the gist of the present invention. Further, in each figure, those having the same reference numerals are the same or equivalent thereof, which are common in the entire text of the specification.

Embodiment 1.
The classification system according to the first embodiment will be described. The configuration of the classification system as shown in FIG. 1 described below is realized by executing the feature analysis processing program read into the auxiliary storage device on a computer (for example, a personal computer or the like). Further, this feature analysis processing program is stored in an information storage medium such as a CD-ROM (Compact Disc-Read Only Memory), or is distributed via a network such as the Internet and installed in a computer.

[Configuration of classification system 1]
The classification system 1 according to the first embodiment will be described. FIG. 1 is a block diagram showing an example of the configuration of the classification system according to the first embodiment. As shown in FIG. 1, the classification system 1 according to the first embodiment includes a learning unit 10 and a classification unit 20. The learning unit 10 and the classification unit 20 are provided in separate devices, for example. Not limited to this, the learning unit 10 and the classification unit 20 may be provided in one device such as an information processing device.

(Learning Department 10)
The learning unit 10 has a time-frequency analysis processing means 11, a feature analysis processing means 12, a learning processing means 13, and a storage means 14. The learning unit 10 inputs the time waveform L101-n (n = 1, ..., N) of N acoustic signals recorded in the past. The time waveform is a waveform showing the amplitude of the time series for the signal to be classified. It is assumed that the time waveform L101-n of the acoustic signal is a signal radiated from any of the known C types of signal sources in advance. Here, the type of this known signal source is referred to as a label, and the time waveform L101-n is represented as a signal of the signal source of the label c (c is any one of 1, ..., C). Indicates the type).

The time frequency analysis processing means 11 receives the time waveform L101-n, performs a time frequency analysis process of converting the time waveform L101-n into data indicating a time series frequency and a signal strength, and performs a time frequency analysis process as a result of the process. Output the result. A typical example of the time-frequency analysis process is a short-time Fourier transform. The time-frequency analysis result in this case is vector data or matrix data.

The feature analysis processing means 12 inputs N time-frequency analysis results and performs feature analysis processing using matrix calculation. Then, the feature analysis processing means 12 generates a feature amount to be supplied to the learning processing means 13, a label of each time waveform L101-n, and a feature extraction parameter to be supplied to the feature extraction processing means 22, and the generated feature amount, Output label and feature extraction parameters.

The input to the feature analysis processing means 12 is an arrangement of N time-frequency analysis results, and becomes matrix data or tensor data. Further, as a typical example of the feature analysis process, there are, for example, non-negative matrix factorization (NMF) or non-negative tensor factorization (NTF). In the first embodiment, the feature analysis processing means 12 performs the feature analysis process by using the submatrix operation as the matrix operation. The details of the feature analysis process using the submatrix operation will be described later.

The learning processing means 13 inputs the feature amount and the label of each time waveform L101-n, estimates the classification parameter to be supplied to the classification processing means 23 from the distribution of the feature amount, and outputs the estimated classification parameter.

The storage means 14 has a main storage area 14a and an auxiliary storage area 14b. The main storage area 14a is an area for temporarily storing necessary data when performing various processes. In the main storage area 14a, for example, when the feature analysis processing means 12 performs a submatrix operation, matrix data or the like necessary for the processing is read from the auxiliary storage area 14b and stored. The auxiliary storage area 14b is an area for storing various types of data, and data is read and written as necessary. The auxiliary storage area 14b stores, for example, matrix data when the feature analysis processing means 12 performs a submatrix operation.

(Classification unit 20)
The classification unit 20 has a time-frequency analysis processing means 21, a feature extraction processing means 22, and a classification processing means 23. The time waveform C101 is input to the classification unit 20. Unlike the time waveform L101-n, the time waveform C101 is unknown from which of the C types of signal sources the signal is radiated. The classification unit 20 estimates the label of this unknown signal.

The time-frequency analysis processing means 21 has the same configuration as the time-frequency analysis processing means 11, performs time-frequency analysis processing on the time waveform C101, and outputs the time-frequency analysis result as the processing result.

The feature extraction processing means 22 refers to the feature extraction parameters supplied from the feature analysis processing means 12, and calculates the feature amount from the time-frequency analysis result. Then, the feature extraction processing means 22 outputs the calculated feature amount.

The classification processing means 23 refers to the classification parameters supplied from the learning processing means 13, and estimates the label c of the unknown time waveform from the feature amount of the unknown time waveform C101. Then, the classification processing means 23 outputs the estimated label C102, which is the estimated label c.

Although not shown, the classification unit 20 is provided with a storage means having a main storage area and an auxiliary storage area, similarly to the learning unit 10.

The learning unit 10 and the classification unit 20 in such a classification system 1 are composed of an arithmetic unit such as a microcomputer that realizes various functions by executing software, or hardware such as a circuit device corresponding to various functions. ing.

FIG. 2 is a hardware configuration diagram showing an example of the configuration of the learning unit of FIG. When various functions of the learning unit 10 are executed by hardware, the learning unit 10 of FIG. 1 is composed of a processing circuit 31 as shown in FIG. In the learning unit 10 of FIG. 1, each function of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13, and the storage means 14 is realized by the processing circuit 31.

When each function is executed by hardware, the processing circuit 31 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), or an FPGA (Field-Programmable Gate). Array), or a combination of these. The learning unit 10 may realize the functions of each part of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13 and the storage means 14 by the processing circuit 31, or the functions of each part may be processed by one. It may be realized by the circuit 31.

FIG. 3 is a hardware configuration diagram showing another example of the configuration of the learning unit of FIG. When various functions of the learning unit 10 are executed by software, the learning unit 10 of FIG. 1 is composed of a processor 41 and a memory 42 as shown in FIG. In the learning unit 10, each function of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13, and the storage means 14 is realized by the processor 41 and the memory 42.

When each function is executed by software, in the learning unit 10, the functions of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13 and the storage means 14 are software, firmware, or software and firmware. It is realized by the combination. The software and firmware are written as a program and stored in the memory 42. The processor 41 realizes the functions of each part by reading and executing the program stored in the memory 42.

As the memory 42, for example, a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM, flash memory, EPROM (Erasable and Programmable ROM) and EEPROM (Electrically Erasable and Programmable ROM) is used. Further, as the memory 42, for example, a removable recording medium such as a magnetic disk, a flexible disk, an optical disk, a CD, an MD (Mini Disc), and a DVD (Digital Versatile Disc) may be used.

In this way, the learning unit 10 can realize each of the above-mentioned functions by hardware, software, firmware, or a combination thereof. As with the learning unit 10, the classification unit 20 can also realize each function of the classification unit 20 by hardware, software, firmware, or a combination thereof.

[Operation of classification system 1]
The operation of the classification system 1 according to the first embodiment will be described. Here, first, in order to facilitate understanding of the operation of the classification system 1 according to the first embodiment, the configuration and operation of the conventional classification system will be described.

(Configuration of conventional classification system 100)
FIG. 4 is a block diagram showing an example of the configuration of a conventional classification system. As shown in FIG. 4, the conventional classification system 100 includes a learning unit 110 and a classification unit 120.

The learning unit 110 has a time-frequency analysis processing means 111, a feature analysis processing means 112, and a learning processing means 113. Similar to the learning unit 10 in the classification system 1 according to the first embodiment, the learning unit 110 has a time waveform L101-n (n = 1, ..., N) of N acoustic signals recorded in the past. Is entered.

Similar to the time frequency analysis processing means 11, the time frequency analysis processing means 111 receives the time waveform L101-n, performs the time frequency analysis processing on the time waveform L101-n, and results in the time frequency analysis as the processing result. Is output.

The feature analysis processing means 112 receives N time-frequency analysis results and performs feature analysis processing using matrix calculation. Then, the feature analysis processing means 112 generates a feature amount to be supplied to the learning processing means 113, a label of each time waveform L101-n, and a feature extraction parameter to be supplied to the feature extraction processing means 122, and the generated feature amount, Output label and feature extraction parameters.

Similar to the learning processing means 13, the learning processing means 113 inputs the feature amount and the label of each time waveform L101-n, and estimates and estimates the classification parameters supplied to the classification processing means 123 from the distribution of the feature amount. Output the classification parameter.

The classification unit 120 has a time-frequency analysis processing means 121, a feature extraction processing means 122, and a classification processing means 123. Similar to the classification unit 20 in the classification system 1 according to the first embodiment, the classification unit 120 inputs the time waveform C101 and estimates the label of the signal based on the input time waveform C101.

The time-frequency analysis processing means 121 has the same configuration as the time-frequency analysis processing means 111, performs time-frequency analysis processing on the time waveform C101, and outputs the time-frequency analysis result as the processing result.

Similar to the feature extraction processing means 22, the feature extraction processing means 122 refers to the feature extraction parameters supplied from the feature analysis processing means 112, and calculates the feature amount from the time-frequency analysis result. Then, the feature extraction processing means 122 outputs the calculated feature amount.

Similar to the classification processing means 23, the classification processing means 123 refers to the classification parameters supplied from the learning processing means 113, and estimates the label c of the unknown time waveform from the feature amount of the unknown time waveform C101. Then, the classification processing means 123 outputs the estimated label C102, which is the estimated label c.

(Operation of the conventional classification system 100)
FIG. 5 is a schematic view showing an example of the result of the time frequency analysis processing by the time frequency analysis processing means. FIG. 6 is a schematic view showing an example of the result of the feature analysis processing by the feature analysis processing means. FIG. 7 is a schematic diagram showing an example of the result of learning processing by the learning processing means. 5 to 7 show images of the time-frequency analysis result 201, the feature extraction parameter 202, the distribution of the feature amount 203, and the classification parameter 204. In FIGS. 5 to 7, it is assumed that the time-frequency analysis result 201 has a plurality of data for each label. Further, FIGS. 5 to 7 illustrate a case where the type of signal source is C = 4.

As shown in FIGS. 5 and 6, the feature extraction parameter 202 is in the form of M vectors and is obtained by decomposing the common frequency components included in the time-frequency analysis result 201. Note that FIG. 6 illustrates a case where the number of vectors of the feature extraction parameter 202 is M = 2.

As for the distribution of the feature amount 203, if the feature extraction parameter 202 is properly obtained, a group is formed for each label as shown in FIG. 7. The number of axes of the distribution of the feature amount 203 is M, which is the same as the number of vectors of the feature extraction parameter 202. The feature amount 203 represents the degree to which the frequency component of the feature extraction parameter 202 is included in each time-frequency analysis result 201.

In FIGS. 5 to 7, the time frequency analysis result 201 of label # 1 will be described as an example. The time-frequency analysis result 201 of label # 1 does not include the first frequency component of the feature extraction parameter 202, but contains the second frequency component. Therefore, looking at the distribution of the feature amount 203 in FIG. 7, as for the feature amount of the label # 1, the value of the feature 1 is small and the value of the feature 2 is large. As shown in FIG. 7, the classification parameter 204 is a parameter representing a boundary line for each label in the distribution of the feature amount 203.

In the conventional classification system 100 that operates in this way, the feature analysis processing means 112 calculates and outputs the feature extraction parameter and the feature amount from the time-frequency analysis result as described above. At this time, the feature analysis processing means 112 performs various calculations on the time-frequency analysis result, which is matrix data or tensor data, by using at least one of the matrix operations of the sum of the matrices and the product of the matrices.

FIG. 8 is a schematic diagram for explaining the sum calculation process of the matrix. As shown in FIG. 8, when calculating the sum of matrices, the sum calculation unit 301 is used. The sum calculation unit 301 adds the two input matrices A and B, and outputs the addition result matrix C (= A + B).

FIG. 9 is a schematic diagram for explaining the matrix multiplication operation process. As shown in FIG. 9, the product calculation unit 302 is used when calculating the product of matrices. The product calculation unit 302 multiplies the two input matrices A and B, and outputs the matrix C (= AB) which is the multiplication result.

(Conventional issue)
By the way, in the matrix calculation by the feature analysis processing means 112, the main storage area and the auxiliary storage area for storing the calculation result are used. The matrix data as the time-frequency analysis result input to the feature analysis processing means 112 is stored in the auxiliary storage area and read into the main storage area during the feature analysis processing. Then, the feature analysis processing means 112 performs the feature analysis process based on the matrix data read into the main storage area, and writes the matrix data obtained by the process into the main storage area.

At this time, in the matrix calculation by the conventional feature analysis processing means 112, at least one of the row size and the column size of the input matrix data is very large, so the following two problems are caused by the insufficient capacity of the main storage area. May occur.

FIG. 10 is a schematic diagram for explaining the first problem caused by the lack of the main storage area. The first problem is that the matrix data stored in the auxiliary storage area cannot be read into the main storage area when the capacity of the main storage area is insufficient. When performing a matrix operation, the matrix data of both the matrix A and the matrix B are read into the main storage area. However, in the example shown in FIG. 10, when the matrix data of the matrix A is read from the auxiliary storage area into the main storage area, the remaining area of the main storage area becomes smaller than the matrix data of the matrix B, and the matrix. The matrix data of B cannot be read into the main storage area.

FIG. 11 is a schematic diagram for explaining the second problem caused by the lack of the main storage area. The second problem is that the matrix data obtained by the matrix operation cannot be read into the main storage area when the main storage area is insufficient. In the example shown in FIG. 11, the matrix data of the matrix A and the matrix B are read from the auxiliary storage area into the main storage area. After that, when the matrix operation is performed using the matrix A and the matrix B and the matrix C is calculated, the remaining capacity of the main storage area becomes smaller than the capacity of the matrix data of the matrix C, and the matrix C Matrix data cannot be read into the main storage area.

As described above, in the conventional classification system 100, at least an area capable of storing both the matrix data used for the matrix operation and the matrix data obtained by the matrix operation is required as the main storage area of the feature analysis processing means 112. Become. That is, in the conventional classification system 100, the main storage area needs to have a capacity of at least three times that of the matrix data.

Therefore, in the feature analysis processing means 12 of the classification system 1 according to the first embodiment, the matrix calculation can be performed even when the capacity of the main storage area 14a is smaller than the capacity conventionally required.

(Matrix calculation by the feature analysis processing means 12 of the classification system 1)
The matrix operation by the feature analysis processing means 12 according to the first embodiment will be described. In the first embodiment, the feature analysis processing means 12 performs the feature analysis process by using the submatrix calculation instead of the conventional matrix calculation.

In the first embodiment, the feature analysis processing means 12 refers to "sum of matrices using submatrix" and "product of matrix using submatrix" with respect to the time-frequency analysis result which is matrix data or tensor data. Various calculations are performed using at least one of the matrix operations.

FIG. 12 is a schematic diagram for explaining the submatrix. _{As shown in FIG. 12, the submatrix is, for example, a M i} × N _j matrix when the M × N matrix A is I-divided in the row direction and J-divided in the column direction. At this time, i = 1, ..., I, and j = 1, ..., J. Further, a sum of _{M i} is M, the sum of _{N j} is N.

(Multiply-accumulate matrix using submatrix)
First, the sum operation of a matrix using a submatrix will be described. FIG. 13 is a schematic diagram for explaining a sum calculation process using a submatrix. As shown in FIG. 13, when the matrix sum calculation process is performed using the submatrix, the first division unit 221 and the plurality of submatrix sum calculation means 222 are used.

The first dividing unit 221, each of the two matrices A and matrix B is input, submatrix _{M i} × _{N j} of the same size _{A ij} and partial matrix _{B ij (i = 1, ···} , I, Divide into j = 1, ..., J). The submatrix sum calculation means 222 is provided according to the number of divided submatrixes. The submatrix sum calculation means 222 _{adds the two input submatrix A ij} and the submatrix B _ij , and outputs the _{submatrix C ij} which is the addition result.

When calculating the sum of matrices using submatrix, the first division unit 221 displays M × N size matrix A and matrix B as I × J, respectively, as shown in equations (1) and (2). Divide into submatrix. In equations (1) and (2), i = 1, ..., I, and j = 1, ..., J. Further, a sum of _{M i} is M, the sum of _{N j} is N. Submatrix _{A ij} and partial matrix _{B ij} are all matrices of the same _{M i} × _{N j} size.

Here, the first division unit 221 holds an index number indicating which part of the matrix A and the matrix B is extracted from the data of the _{submatrix A ij} and the submatrix B _{ij, and the matrix A and the matrix B.} Not all the data of B is read into the main storage area 14a at once.

Next, the submatrix sum calculation means 222 takes the submatrix A _ij and the submatrix B _ij as inputs, and outputs the _{submatrix C ij.} The correspondence between input and output in this case is shown in the circled portion of FIG.

FIG. 14 is a block diagram showing an example of the configuration of the submatrix sum calculation means of FIG. As shown in FIG. 14, the submatrix sum calculation means 222 includes a sum calculation unit 223. In this way, when calculating the sum of the submatrix, the sum calculation unit 223 is used. The sum calculation unit 223 _{adds the two input submatrix A ij} and the submatrix B _ij , and outputs the _{submatrix C ij} which is the addition result.

FIG. 15 is a schematic view showing an example of the state of the main storage area and the auxiliary storage area of the sum operation processing using the submatrix. At this time, as shown in FIG. 15, the feature analysis processing means 12 _{reads the submatrix Aij} and the submatrix _Bij from the auxiliary storage area 14b into the main storage area 14a, and uses the equation (3) to read the submatrix _Cij. To calculate. Then, the feature analysis processing means 12 _{writes the submatrix Cij} , which is the calculation result, from the main storage area 14a to the auxiliary storage area 14b.

Finally, when the submatrix _Cij is calculated for all i and j, the sum of the matrices C = A + B is obtained, as shown in equation (4).

(Matrix multiplication operation using submatrix)
Next, a matrix multiplication operation using a submatrix will be described. FIG. 16 is a schematic diagram for explaining a product calculation process using a submatrix. As shown in FIG. 16, when the matrix product calculation process is performed using the submatrix, the second division unit 224 and the plurality of submatrix product calculation means 225 are used.

An M × L size matrix A and an L × N size matrix B are input to the second division unit 224. The second dividing unit 224, a matrix A of the input _{M × L, M i × L} k size submatrix _{A ik (i = 1, ···} , I, k = 1, ···, K) Divide into. Further, the second division unit 224 uses the input L × N matrix B as a submatrix B _kj (k = 1, ..., K, j = 1, ..., Of _{L k} × N _{j size).} Divide into J). The submatrix product calculation means 225 calculates the submatrix C _{ij based on the two input submatrix A ik} and the submatrix B _jk .

When calculating the product of a matrix using a submatrix, the second division unit 224 divides the M × L size matrix A into I × K submatrixes as shown in the equation (5). Further, the second division unit 224 divides the L × N size matrix B into K × J submatrixes as shown in the equation (6). In equations (5) and (6), i = 1, ..., I, j = 1, ..., J, and k = 1, ..., K. Further, a sum of _{M i} is M, is the sum of _{N j} is N, the sum of _{L k} is L. Furthermore, the partial matrix _{A ik} is a matrix of _{M i} × _{L k} size, partial matrix _{B kj} is a matrix of _{L k} × _{N j} size.

Here, the second division unit 224 holds an index number indicating which part of the matrix A and the matrix B is extracted from the data of the _{submatrix A ik} and the submatrix B _{kj, and the matrix A and the matrix B.} Not all the data of B is read into the main storage area 14a at once.

Next, the submatrix product calculation means 225 takes the submatrix A _ik and the submatrix B _jk as inputs, and outputs the _{submatrix C ij.} The correspondence between input and output in this case is shown in the circled portion of FIG. The submatrix product calculation means 225 calculates the submatrix C _{ij from the input submatrix A ik} and the submatrix B _jk using the equation (7).

FIG. 17 is a block diagram showing an example of the configuration of the partial matrix product calculation means of FIG. FIG. 18 is a schematic view showing an example of the state of the main storage area and the auxiliary storage area of the product calculation processing using the submatrix. As shown in FIG. 17, the partial matrix product calculation means 225 includes a product calculation unit 226 and a sum calculation unit 227. In this way, when calculating the product of sub-matrix, the product calculation unit 226 and the sum calculation unit 227 are used.

The product calculation unit 226 multiplies the two input submatrix A _ij and the submatrix B _ij , and outputs the _{submatrix A ij} B _ij which is the multiplication result. The sum calculation unit 227 adds the two input submatrixes and outputs the _{submatrix Cij which is the addition result.}

In this case, as shown in FIG. 18, first, the submatrix A _ik and the submatrix B _jk are read from the auxiliary storage area 14b into the main storage area 14a, and the submatrix C _ij is calculated based on the equation (7). Then, the calculated submatrix _Cij is written from the main storage area 14a to the auxiliary storage area 14b.

In the actual calculation by the submatrix product calculation means 225, first, the submatrix product calculation means 225 _{initializes the output submatrix C ij} with the zero matrix O. Next, the submatrix product calculation means 225 _{reads the submatrix A ik} and the submatrix B _kj from the auxiliary storage area 14b into the main storage area 14a _{as the first product calculation process, and the product calculation unit 226 reads A ik} B _kj. Is calculated. _{Further, the submatrix product calculation means 225 calculates "C ij} + A _ik B _kj " in the sum calculation unit 227 as the second product calculation process, and overwrites the _{submatrix C ij} read in the main storage area 14a. Update by doing. That is, the submatrix product calculation means 225 accumulates the calculated _{A ik} B _kj with respect to the _{submatrix C ij read in the main storage area 14a.}

The submatrix product calculation means 225 repeats the first and second product calculation processes for all k, _{and writes the finally obtained submatrix Cij} from the main storage area 14a to the auxiliary storage area 14b. As a result, the operation corresponding to the equation (7) is performed.

_{Finally, when the calculation of the submatrix Cij} for all i and j is completed, the submatrix product calculation means 225 calculates the matrix product "C = AB" as shown in the equation (8).

As described above, in the first embodiment, when the submatrix is used in the matrix sum operation, the matrix C = A + B is output as in the conventional matrix sum operation. Further, when a submatrix is used in the matrix multiplication operation, the matrix C = AB is output as in the conventional matrix multiplication operation.

Further, in the first embodiment, if the capacity of the main storage area 14a is at least about three times the size of the submatrix used for the calculation, the matrix C can be obtained by the sum calculation of the matrix using the submatrix. can. Further, if the capacity of the main storage area 14a is at least about 4 times the size of the submatrix used for the calculation, the matrix C can be obtained by the matrix multiplication operation using the submatrix.

As a result, it is possible to improve the problem that the matrix data cannot be read due to the lack of the main storage area or the matrix data of the calculation result cannot be held in the main storage area, which may occur in the prior art. Then, since it becomes possible to process a huge matrix data, the classification ability by the classification unit 20 can be improved.

As described above, in the classification system 1 according to the first embodiment, the submatrix operation is used in the feature analysis process. Therefore, for the purpose of improving the classification ability of the classification unit 20, even when the learning unit 10 handles a huge matrix data, the main storage area may be provided according to the size of the submatrix. Even when the capacity is suppressed, the matrix calculation can be performed appropriately.

1 Classification system, 10 learning unit, 11-hour frequency analysis processing means, 12 feature analysis processing means, 13 learning processing means, 14 storage means, 14a main storage area, 14b auxiliary storage area, 20 classification unit, 21-hour frequency analysis processing means , 22 Feature extraction processing means, 23 Classification processing means, 31 Processing circuit, 41 Processor, 42 Memory, 100 Classification system, 110 Learning unit, 111 Time frequency analysis processing means, 112 Feature analysis processing means, 113 Learning processing means, 120 classification Unit, 121 hour frequency analysis processing means, 122 feature extraction processing means, 123 classification processing means, 201 hour frequency analysis result, 202 feature extraction parameter, 203 feature quantity, 204 classification parameter, 221 first division part, 222 partial matrix sum calculation Means, 223, 227, 301 Sum calculation unit, 224 Second division unit, 225 Partial matrix product calculation means, 226, 302 Product calculation unit.

Claims (6)

A classification system that performs feature extraction and classification based on time-frequency analysis.
A time-frequency analysis processing means for converting a time waveform indicating the time-series amplitude of a known signal whose signal source is known into data indicating the time-series frequency and signal strength.
A feature analysis processing means for generating feature quantities and feature extraction parameters using submatrix operations on the data, and
A learning unit having a learning processing means for estimating classification parameters from the feature amount, and
With reference to the feature extraction parameters, a feature extraction processing means for calculating the feature amount of an unknown signal whose signal source is unknown, and
A classification system including a classification unit having a classification processing means for estimating a signal source of the unknown signal from the feature amount of the unknown signal with reference to the classification parameter. The feature analysis processing means is
The classification system according to claim 1, wherein the sum of matrices using a submatrix is calculated as the submatrix operation. The feature analysis processing means is
The classification system according to claim 1 or 2, wherein as the submatrix operation, the product of matrices using a submatrix is calculated. An information processing device that performs feature extraction and classification based on time-frequency analysis.
A time-frequency analysis processing means for converting a time waveform indicating the time-series amplitude of a known signal whose signal source is known into data indicating the time-series frequency and signal strength.
A feature analysis processing means for generating feature quantities and feature extraction parameters using submatrix operations on the data, and
A learning processing means for estimating classification parameters from the features, and
With reference to the feature extraction parameters, a feature extraction processing means for calculating the feature amount of an unknown signal whose signal source is unknown, and
An information processing device including a classification processing means for estimating a signal source of the unknown signal from the feature amount of the unknown signal with reference to the classification parameter. It is a classification method that performs feature extraction and classification based on time-frequency analysis.
A time waveform showing the time-series amplitude of a known signal whose signal source is known is converted into data showing the time-series frequency and signal strength.
Feature quantities and feature extraction parameters are generated for the data using submatrix operations.
The classification parameters are estimated from the features, and the classification parameters are estimated.
With reference to the feature extraction parameters, the feature amount of an unknown signal whose signal source is unknown is calculated.
A classification method for estimating the signal source of the unknown signal from the feature amount of the unknown signal with reference to the classification parameter. Computers that perform feature extraction and classification based on time-frequency analysis
A time-frequency analysis processing means for converting a time waveform indicating the time-series amplitude of a known signal whose signal source is known into data indicating the time-series frequency and signal strength.
A feature analysis processing means for generating feature quantities and feature extraction parameters using submatrix operations on the data.
Learning processing means for estimating classification parameters from the features,
A feature extraction processing means for calculating a feature amount of an unknown signal whose signal source is unknown with reference to the feature extraction parameter.
A program for functioning as a classification processing means for estimating a signal source of the unknown signal from the feature amount of the unknown signal with reference to the classification parameter.

Images