JP6663765B2 - State determination device, program, and integrated circuit - Google Patents

Description

  The present invention relates to a technique for analyzing input data using a self-organizing map to determine a state (pattern) indicated by the input data.

  The self-organizing map (hereinafter, referred to as “SOM”) models the visual cortex of the cerebral cortex. In SOM, multidimensional data can be visualized (for example, pattern classification) by mapping input data to an arbitrary dimension by unsupervised learning. For example, in SOM, high-dimensional input data can be projected onto a two-dimensional map by mapping high-dimensional input data to, for example, two-dimensional data. By using the two-dimensional data (two-dimensional map) thus projected, a classification process (a pattern classification process) of the input data can be performed. In the SOM, by observing the two-dimensional data (two-dimensional map) projected as described above, it is easy to visually grasp the classification state of the input data. Therefore, the SOM is used in various fields. Used.

  For example, Patent Literature 1 discloses a technique for performing waveform analysis using SOM.

In the technique of Patent Literature 1, a waveform map is generated using SOM such that an area representing a predetermined waveform pattern is a continuous area. In the technique of Patent Literature 1, as shown in FIG. 16 of Patent Literature 1, the image areas on the following waveform map are each made to be a continuous image area (without a divided image area) so as to be a clear image area. Generate a segmented waveform map.
(1) Image area showing waveform A + (2) Image area showing waveform A (3) Image area showing waveform B + (4) Image area showing waveform B (5) Image area showing waveform C + (6) Waveform C (7) Image area showing waveform D + (8) Image area showing waveform D In the technique of Patent Document 1, input processing is performed by performing a waveform pattern classification process using the above-described waveform map. The obtained waveform data can be appropriately classified (determined to be similar to a predetermined waveform pattern).

JP 2007-202964 A

  However, in the technique of Patent Literature 1, it is necessary to generate a waveform map using SOM such that an area representing a predetermined waveform pattern is a continuous area, and a discontinuous image area (for example, a divided image area) It is difficult to perform a classification process and a pattern determination process using a waveform map having a region. That is, in the related art, if a discontinuous area (for example, a divided area) exists in the map generated by the SOM, it is difficult to appropriately perform the pattern classification processing and the pattern determination processing.

  In view of the above problems, the present invention provides a method for appropriately classifying a pattern even if a discontinuous area (for example, a divided area) exists in a map generated by SOM. It is an object of the present invention to realize a state determination device, a program, and an integrated circuit that can perform a determination process.

  In order to solve the above-described problems, a first invention is a state determination device including a feature vector acquisition unit, a mapping conversion unit, a matching processing unit, and a state determination unit.

  The feature vector acquisition unit acquires feature vector data from measurement data obtained by measuring an event in an unknown state.

  The mapping conversion unit acquires the SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data.

  The matching processing unit acquires matching degree data indicating a degree of correlation between the template data indicating a predetermined state and the SOM output data acquired by the mapping conversion unit.

  The state determination unit determines a state indicated by the measurement data based on the fitness data acquired by the matching processing unit.

  In this state determination device, the matching processing unit acquires the conformity data indicating the degree of correlation between the template data indicating the predetermined state and the SOM output data, and the state determination unit determines the input data based on the conformity data. Is determined.

  That is, in the state determination device, even when a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the template indicating the predetermined state and the SOM output The degree of correlation (similarity) with data can be detected. Therefore, in the state determination device, even when a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the pattern classification processing and the pattern determination processing are appropriately performed. be able to.

  In a second aspect based on the first aspect, the matching processing section determines a correlation between each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data. The acquired conformity data D_f (1) to D_f (M) are obtained.

  The state determination unit specifies template data corresponding to the fitness data indicating the highest correlation among the fitness data D_f (1) to D_f (M), and determines a state corresponding to the identified template data. It is determined that the state is indicated by the measurement data.

  This allows the state determination device to recognize the degree of correlation between the SOM output data acquired from the input data and the M template data Tmpl (1) to Tmpl (M) from the matching degree data. Therefore, this state determination device can appropriately determine which of the M states corresponding to the M template data Tmpl (1) to Tmpl (M) the input data is close to.

  That is, in the state determination device, even when a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the template indicating the predetermined state and the SOM output The degree of correlation (similarity) with data can be detected. Therefore, in the state determination device, even when a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the pattern classification processing (the M states) is appropriately performed. Pattern classification processing) and pattern determination processing.

  A third invention is the second invention, wherein the matching processing unit calculates an inner product of each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data. Then, the value of the calculated inner product is acquired as the fitness data D_f (1) to D_f (M).

  The state determination unit specifies template data corresponding to the maximum degree of fitness data among the fitness data D_f (1) to D_f (M), and the measurement data indicates a state corresponding to the specified template data. It is determined that it is in the state.

  Thus, in this state determination device, the degree of correlation between the SOM output data obtained from the input data and each of the M template data Tmpl (1) to Tmpl (M) is obtained by calculating the inner product. Can be.

  In a fourth aspect based on the second aspect, the matching processing section determines a Euclidean distance between each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data. The calculated Euclidean distance value is obtained as the conformity data D_f (1) to D_f (M).

  The state determination unit specifies template data corresponding to the fitness data having the minimum value among the fitness data D_f (1) to D_f (M), and the measurement data indicates a state corresponding to the identified template data. It is determined that it is in the state.

  Thus, in this state determination device, the degree of correlation between the SOM output data obtained from the input data and each of the M template data Tmpl (1) to Tmpl (M) is obtained by calculating the Euclidean distance. be able to.

  In a fifth aspect based on any one of the first to fourth aspects, the template data is data composed of two-dimensional elements, and the sum of the values of each element data of the template data is “1”. ".

  In this state determination device, the sum of the values of the data of each element of the template data is normalized to be “1”. Thus, in the state determination device, in the process of determining the degree of correlation between the template data and the SOM output data, a value indicating the degree of correlation (for example, the value of the inner product or the Euclidean distance) is used regardless of which template is used. ) Can be constant (common), which facilitates the classification determination process.

  A sixth invention is any one of the first to fourth inventions, wherein the template data is data composed of two-dimensional elements.

  The template data is data generated based on the N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) where the same state continues.

When the template data is data composed of m × n (m, n: natural numbers), the two-dimensional coordinates (i, j) of the template data (i, j: natural numbers, 1 ≦ i ≦ m, 1 ≦ j ≦ n) is _defined as g _ij, and the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is defined as D _ij (t).
The template data is
Generated by _converting the value g _ij acquired by the above into element data of the two-dimensional coordinates (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) of the template. It is.

  This allows the state determination device to execute the state determination process using the template obtained by averaging the SOM output data acquired during the period in which the same state continues for the time axis direction.

  The seventh invention is any of the first to fifth inventions, wherein the template data is data composed of two-dimensional elements.

  The template data is data generated based on the N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) where the same state continues.

When the template data is data composed of m × n (m, n: natural numbers), the two-dimensional coordinates (i, j) of the template data (i, j: natural numbers, 1 ≦ i ≦ m, 1 ≦ j ≦ n) is _defined as g _ij, and the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is defined as D _ij (t).
The template data compares the element data D _ij (t) with a predetermined threshold Th1, calculates an average value of the element data D _ij (t) that satisfies D _ij (t) ≧ Th1, and coordinates the calculated average value with coordinates ( The data is generated by setting the value g _ij of the template data of (i, j).

  This allows the state determination device to execute the state determination process using the template obtained by averaging the SOM output data acquired during the period in which the same state continues for the time axis direction. Since the template data is data generated by excluding data smaller than a predetermined threshold from the target of the averaging process, by performing processing using such template data, in this state determination device, Appropriate state determination processing can be performed.

  The eighth invention is any of the first to fifth inventions, wherein the template data is data composed of two-dimensional elements.

  The template data is data generated based on the N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) where the same state continues.

When the template data is data composed of m × n (m, n: natural numbers), the two-dimensional coordinates (i, j) of the template data (i, j: natural numbers, 1 ≦ i ≦ m, 1 ≦ j ≦ n) is _defined as g _ij, and the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is defined as D _ij (t).
The template data is
The variance value v _ij of the element data at the coordinates (i, j) obtained by the above is compared with a predetermined threshold Th2, and the value D _ij of the SOM output data at the coordinates (i, j) satisfying v _ij ≤ Th2 This is data generated by calculating the average value of t) and using the calculated average value as the value g _ij of the template data at the coordinates (i, j).

  This allows the state determination device to execute the state determination process using the template obtained by averaging the SOM output data acquired during the period in which the same state continues for the time axis direction. Then, since the template data is data obtained by averaging processing only data having a variance in the time axis direction smaller than a predetermined threshold, by processing using such template data, In this state determination device, an appropriate state determination process can be performed.

  A ninth invention is the invention according to any one of the first to fifth inventions, wherein the template data is data composed of two-dimensional elements.

  The template data is data generated based on the N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) where the same state continues.

When the template data is data composed of m × n (m, n: natural numbers), the two-dimensional coordinates (i, j) of the template data (i, j: natural numbers, 1 ≦ i ≦ m, 1 ≦ j ≦ n) is _defined as g _ij, and the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is defined as D _ij (t).
The template data is
The variance value v _ij of the element data at the coordinates (i, j) obtained by the above is compared with a predetermined threshold Th2, and the value D _ij of the SOM output data at the coordinates (i, j) satisfying v _ij ≤ Th2 t), and calculate the average value of only the SOM output data in which the value D _ij (t) is greater than a predetermined threshold Th3, and use the calculated average value as template data of coordinates (i, j). _Is the data generated by setting the value g _{ij of}

  This allows the state determination device to execute the state determination process using the template obtained by averaging the SOM output data acquired during the period in which the same state continues for the time axis direction. Then, the template data is data obtained by averaging processing only element data having a variance in the time axis direction smaller than a predetermined threshold value and having a value larger than the predetermined threshold value. Therefore, by performing processing using such template data, this state determination device can execute appropriate state determination processing.

  A tenth invention is the invention according to any one of the first to fifth inventions, wherein the template data is data composed of two-dimensional elements.

The template data is data generated based on N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) in which the same state continues. , M × n (m, n: natural numbers), two-dimensional coordinates (i, j) of template data (i, j: natural numbers, 1 ≦ i ≦ m, 1 ≦ j ≦ If the element data of n) is g _ij and the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is D _ij (t),
The template data is
σ _ij = sqrt (v _ij )
sqrt (x): When the standard deviation of the element data at the coordinates (i, j) obtained by the function for obtaining the square root of x is σ _ij ,
When abs (D _ij (t) −AveD _ij ) ≦ n1 × σ _ij (n1: a positive real number, abs (x) is a function for obtaining the absolute value of x), the SOM output of the coordinates (i, j) is satisfied. of an average value of the values _D ij (t), and the calculated average value coordinates (i, j) and the value _{g ij} of the template data,
If abs (D _ij (t) −AveD _ij ) ≦ n1 × σ _ij is not satisfied, the data is generated by setting the value g _ij of the template data at the coordinates (i, j) to “0”.

  This allows the state determination device to execute the state determination process using the template obtained by averaging the SOM output data acquired during the period in which the same state continues for the time axis direction. Then, the template data is data obtained by averaging processing on data values in which the value of each element data is within 3σ from the average value, and thus processing is performed using such template data. This allows the state determination device to execute an appropriate state determination process.

  An eleventh invention is a program for causing a computer to execute a state determination method.

  The state determination method includes a feature vector acquisition step, a mapping conversion step, a matching processing step, and a state determination step.

  The feature vector obtaining step obtains feature vector data from measurement data obtained by measuring an event in an unknown state.

  The mapping transformation step acquires SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data.

  The matching processing step acquires conformity data indicating the degree of correlation between the template data indicating a predetermined state and the SOM output data acquired in the mapping transformation step.

  The state determination step determines a state indicated by the measurement data based on the matching degree data acquired in the matching processing step.

  Accordingly, it is possible to realize a program for causing a computer to execute a state determination method having the same effect as the first invention.

  A twelfth invention is an integrated circuit including a feature vector acquisition unit, a mapping conversion unit, a matching processing unit, and a state determination unit.

  The feature vector acquisition unit acquires feature vector data from measurement data obtained by measuring an event in an unknown state.

  The mapping conversion unit acquires the SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data.

  The matching processing unit acquires matching degree data indicating a degree of correlation between the template data indicating a predetermined state and the SOM output data acquired by the mapping conversion unit.

  The state determination unit determines a state indicated by the measurement data based on the fitness data acquired by the matching processing unit.

  Thus, an integrated circuit having the same effects as the first invention can be realized.

  Advantageous Effects of Invention According to the present invention, even when a discontinuous area (for example, a divided area) exists in a map generated by SOM, it is possible to appropriately perform pattern classification processing and pattern determination processing. A state determination device, a program, and an integrated circuit that can be realized can be realized.

FIG. 2 is a schematic configuration diagram of a state determination device 1000 according to the first embodiment. FIG. 2 is a schematic configuration diagram of a matching processing unit 8 according to the first embodiment. FIG. 4 is a schematic configuration diagram of a matching processing unit 8 when M = 3. FIG. 3 is a diagram showing digital data D1 (time-series data D1) obtained by the data input unit 1, and frequency domain data D2 obtained by frequency-converting the digital data D1. FIG. 4 is a diagram for explaining the SOM process, and is a diagram schematically showing the structure of the SOM. The figure which shows an example of template data. 11 is a flowchart of processing 1 (output value threshold processing). FIG. 5 is a diagram illustrating a waveform of a signal D1 (waveform in a walking state) acquired by the data input unit 1, a feature vector acquiring unit 2 acquiring feature vector data D2 acquired from the signal D1, and a mapping transformation unit 7 acquiring a feature vector The figure which shows SOM output data D_som (1) acquired from data D2. FIG. 6 is a diagram showing a waveform of a signal D1 (waveform in a Running state) acquired by the data input unit 1, a feature vector acquisition unit 2 acquiring feature vector data D2 acquired from the signal D1, and a mapping conversion unit 7 acquiring a feature vector The figure which shows SOM output data D_som (2) acquired from data D2. The schematic block diagram of the state determination apparatus 1000A which concerns on the 1st modification of 1st Embodiment. FIG. 9 is a schematic configuration diagram of a matching processing unit 8A according to a first modification of the first embodiment. FIG. 2 is a diagram showing a CPU bus configuration.

[First Embodiment]
The first embodiment will be described below with reference to the drawings.

<1.1: Configuration of state determination device>
FIG. 1 is a schematic configuration diagram of a state determination device 1000 according to the first embodiment.

  FIG. 2 is a schematic configuration diagram of the matching processing unit 8 of the first embodiment.

  As shown in FIG. 1, the state determination device 1000 includes a data input unit 1, a feature vector acquisition unit 2, a switch SEL1, an SOM processing unit 3, an SOM data storage unit 4, a template data generation unit 5, , A template data storage unit 6.

  Further, the state determination device 1000 includes a mapping conversion unit 7, a matching processing unit 8, and a state determination unit 9, as shown in FIG.

  The data input unit 1 receives, for example, an electric signal (sensor output signal) output from a center (for example, an acceleration sensor). The data input unit 1 converts the input electric signal into a digital signal, and outputs the converted signal to the feature vector acquisition unit 2 as a signal D1.

  The feature vector acquisition unit 2 receives the signal D1 output from the data input unit 1. The feature vector acquisition unit 2 performs, for example, frequency conversion on the input signal D1, and acquires feature vector data. Then, the feature vector acquisition unit 2 outputs the acquired feature vector data to the switch SEL1 as a signal D2 (feature vector data D2).

  The switch SEL1 receives the signal D2 output from the feature vector acquisition unit 2. The switch SEL1 outputs the signal D2 to the SOM processing unit 3 or the mapping conversion unit 7 according to the selection signal sel1. Note that the selection signal sel1 is a control signal generated by a control unit (not shown) that controls each functional unit of the state determination device 1000.

The SOM processing unit 3 inputs the feature vector data D2 (signal D2) output from the switch SEL1. The SOM processing unit 3 performs an SOM process using the feature vector data D2.
(1) In a learning phase (a phase in which a process of determining a coupling coefficient between each node in the input layer and each neuron in the output layer) is performed, the SOM processing unit 3 executes the SOM process using the feature vector data D2. Then, a coupling coefficient between each node in the input layer and each neuron in the output layer is determined. The SOM processing unit 3 outputs the data of the coupling coefficient between each node of the input layer and the determined neuron of the output layer, which has been completed after the above processing, to the mapping conversion unit 7 as coupling coefficient data w _{ij —} fixed.
(2) In the template creation phase, the SOM processing unit 3 executes the SOM processing using the feature vector data D2, and converts the SOM output data into data D_som (t) (SOM output data D_som (t) at the current time t). ) Is output to the SOM data storage unit 4.

  The SOM data storage unit 4 stores SOM output data D_som (t) output from the SOM processing unit 3. The SOM data storage unit 4 is realized by, for example, a RAM (Random Access Memory).

  The template data generating unit 5 outputs a plurality of SOM output data D_som from the SOM data storage unit 4, for example, N (N: natural number) SOM output data D_som (t) to N obtained from time t to time t + N-1. Read D_som (t + N-1). The plurality of SOM data read from the SOM data storage unit 4 by the template data generation unit 5 is referred to as vector data vec_D_som. The template data generation unit 5 generates template data from the plurality of SOM data read from the SOM data storage unit 4, and outputs the generated template data to the template data storage unit 6 as template data Tmpl.

  The template data storage unit 6 stores the template data Tmpl output from the template data generation unit 5. The template data storage unit 6 is realized by, for example, a RAM (Random Access Memory).

The mapping conversion unit 7 receives the feature vector data D2 (signal D2) output from the switch SEL1. Also, mapping transform unit 7, the SOM processor 3 inputs are determined, and each node of the input layer, the coupling coefficient data _w ij _fixed with each neuron in the output layer. The mapping conversion unit 7 performs SOM processing on the feature vector data D2 (signal D2) using the coupling coefficient data w _{ij —} fixed to obtain SOM output data D_som. Then, the mapping conversion unit 7 outputs the obtained SOM output data D_som to the matching processing unit 8.

  The matching processing unit 8 receives the SOM output data D_som output from the mapping conversion unit 7. Further, the matching processing unit 8 reads a plurality of template data Tmpl from the template data storage unit 6. The plurality of template data read from the template data storage unit 6 by the matching processing unit 8 is referred to as vector data vec_Tmpl.

  The matching processing unit 8 performs a matching process using the SOM output data D_som and a plurality of template data read from the template data storage unit 6, and acquires the conformity data D_f for each template data. Then, the matching processing unit 8 outputs the acquired fitness data D_f for each template data to the state determination unit 9 as vector data vec_D_f.

  As shown in FIG. 2, the matching processing unit 8 includes a template data acquisition unit 81 and M (M: natural number) inner product calculation units 821 to M which are M (natural numbers) inner product calculation units. , Is provided.

  The template data acquisition unit 81 reads a plurality of template data Tmpl (M template data Tmpl (1) to Tmpl (M)) from the template data storage unit 6. Then, the template data obtaining unit 81 outputs the read M pieces of template data Tmpl (1) to Tmpl (M) to the first inner product calculating unit 821 to the Mth inner product calculating unit 82M, respectively.

  The first inner product calculation unit 821 receives the SOM output data D_som output from the mapping conversion unit 7 and the template data Tmpl (1) output from the template data acquisition unit 81. The first inner product calculation unit 821 executes a process of calculating an inner product using the SOM output data D_som and the template data Tmpl (1). Then, the first inner product calculation unit 821 outputs a signal indicating the calculated value of the inner product to the state determination unit 9 as a signal D_f (1).

  The same applies to the second inner product calculating unit 822 to the M-th inner product calculating unit 82M.

  That is, the k-th inner product calculation unit 82k receives the SOM output data D_som output from the mapping conversion unit 7 and the template data Tmpl (k) output from the template data acquisition unit 81 (k: natural number; 1). ≤ k ≤ M). The k-th inner product calculation unit 82k executes a process of calculating an inner product using the SOM output data D_som and the template data Tmpl (k). Then, the k-th inner product calculation unit 82k outputs a signal indicating the calculated value of the inner product to the state determination unit 9 as a signal D_f (k).

  The signals D_f (1) to D_f (M) output from the first inner product calculating unit 821 to the M-th inner product calculating unit 82M are collectively referred to as vector data vec_Tmpl.

  In the following, a case where M = 3 will be described for convenience of explanation.

  FIG. 3 is a schematic configuration diagram of the matching processing unit 8 when M = 3.

  The state determination unit 9 receives the vector data vec_D_f of the goodness-of-fit data output from the matching processing unit 8. The state determination unit 9 determines the state at the current time t based on the input vector data vec_D_f, and outputs data including the determination result to the outside as data Dout.

<1.2: Operation of state determination device>
The operation of the state determination device 1000 configured as described above will be described below.

  Hereinafter, an example in which an output from an acceleration sensor (not shown) for detecting the state of a person is input to the state determination device 1000 will be described. For convenience of explanation, the human state includes (1) a walking state (walking state), (2) a running state (running state), and (3) a standing state (standing state). The case where three states (states of human actions) are determined by the state determination device 1000 will be described below.

  Hereinafter, the operation of the state determination device 1000 will be described separately for (1) the operation of the learning phase, (2) the operation of the template creation phase, and (3) the operation of the state determination processing phase.

(1.2.1: Learning phase operation)
First, the operation of the learning phase of the state determination device 1000 will be described.

  An output from an acceleration sensor (not shown) for detecting the state of a person is input to the data input unit 1 as a signal Din. The acceleration sensor is mounted on, for example, a human body.

  The data input unit 1 samples (A / D converts) the signal Din and obtains it as discrete data (digital data). Then, the data input unit 1 outputs the obtained discrete data (digital data) to the feature vector obtaining unit 2 as a signal D1.

  The feature vector acquisition unit 2 converts each of the digital data D1 by converting the discrete data D1 (digital data D1) acquired by the data input unit 1 into frequency domain data using, for example, Gabor wavelet transform. Acquire the intensity and phase of the frequency spectrum.

  Then, the feature vector acquisition unit 2 acquires, for example, data of a part of the band of the frequency spectrum intensity acquired by the above processing as the feature vector data D2.

  Note that the transform used to acquire frequency domain data from the discrete data D1 (digital data D1) is not limited to the Gabor wavelet transform. For example, a discrete Fourier transform, a fast Fourier transform, a wavelet transform, etc. May be used.

  FIG. 4 shows digital data D1 (time-series data D1) acquired by the data input unit 1 (upper figure) and frequency domain data D2 (lower figure) acquired by frequency-converting the digital data D1. FIG.

  In the case of FIG. 4, the feature vector acquisition unit 2 uses frequency domain data D2 as eight-dimensional data (vector data including eight parameters (data of eight frequency components included in a range of 0 to 5 Hz)). Have acquired.

  In the following, for convenience of explanation, the feature vector acquisition unit 2 converts the eight-dimensional data (vector data including eight parameters (data of eight frequency components included in a range of 0 to 5 Hz)) into data D2. It will be described as what is acquired as.

  As described above, the feature vector data D2 acquired by the feature vector acquiring unit 2 is output from the feature vector acquiring unit 2 to the switch SEL1.

  A control unit (not shown) that controls each functional unit of the state determination device 1000 generates a selection signal sel1 for selecting the terminal 0 of the switch SEL1, and outputs the generated selection signal sel1 to the switch SEL1.

  The switch SEL1 outputs a signal D2 to the SOM processing unit 3 according to the selection signal sel1.

  The SOM processing unit 3 performs SOM processing using the feature vector data D2 (signal D2) output from the switch SEL1, and determines a coupling coefficient between each node of the input layer and each neuron of the output layer. . This processing will be described with reference to FIG.

  FIG. 5 is a diagram for explaining the SOM process, and is a diagram schematically illustrating the structure of the SOM. As shown in FIG. 5, an i-axis and a j-axis for defining a plane on which the neurons of the output layer are arranged are set.

Input data is given by an n-dimensional real vector vec_x = (x ₁ , x ₂ ,..., X _n ), and a two-dimensional SOM has neurons arranged on m1 × m2 (= M1) grid points. And That is, as shown in FIG. 5, the input layer has n nodes, and the output layer has neurons arranged on m1 × m2 (= M1) grid points.

Each node (each of n input nodes) of the input layer for inputting the input data vec_x is connected to all neurons (all output nodes). The neuron ui _{, j} located at (i, j) on the two-dimensional lattice array of the output layer has a variable connection weight vector vec_w _ij = () for adding a weight to each element of the input data vec_x and connecting the elements. w _{ij — 1} , w _{ij — 2} ,..., w _{ij — n} ). The coupling weight vector Vec_w _ij is referred to as a reference vector.

Note that w _{ij —} k is a connection weight between the output node at the coordinates (i, j) and the k-th input node.

  The SOM processing unit 3 executes the SOM processing by the following {Step 1} and {Step 2} processing.

{Step 1: Initialization}
Assuming that the number of input data vectors is n and the number of repetitions is T (> = n), the SOM processing unit 3 randomly sets initial values of vec_w ₁₁ , vec_w ₁₂ ,..., Vec_w _mm at time t = 0. (Set using random numbers).

≪Step 2: Learning≫
At time t, the SOM processing unit 3 performs the following operation.
(2A):
SOM processor 3, an input data Vec_x of time t (t), the Euclidean distance dis between each reference vector _{Vec_w} ij at time t-1 (t-1) (vec_x (t), vec_w ij (t- 1)) is determined. Note that dis (vec1, vec2) is a function for calculating the Euclidean distance between the vector vec1 and the vector vec2.
(2B):
The SOM processing unit 3 specifies the neuron u _IJ having the minimum Euclidean distance, and sets the specified neuron as the winning neuron u _IJ . Neuron u _IJ is a neuron at coordinates (I, J).

The SOM processing unit 3 specifies the input data vec_x (t) at time t, and the reference vector vec_w _ij (t-1) at time t−1, and specifies the neuron u _IJ whose inner product has the maximum value. The resulting neuron may be the winning neuron u _IJ .
(2C):
SOM processor 3, the reference vector _{Vec_w} ij at time t (t), is learned by the following equation (acquisition).
_{vec_w ij (t) = vec_w ij} (t-1) + h ((i, j), (I, J), t) × (vec_x (t) -vec_w ij (t-1))
Here, h is a function called a neighborhood function, and is a function having the following properties.
(1) The function h is a monotonically decreasing function related to t, and converges to “0” when t is infinite.
(2) The function h is a monotonically decreasing function with respect to the Euclidean distance dis (vec_u _ij , vec_u _IJ ) between the grid point (i, j) and the grid point (I, J). Further, the degree of the monotone decrease of the function h increases as the value of t increases. Note that vec_u _ij is a vector from a predetermined point (for example, the origin) to the position of the neuron u _ij , and vec_u _IJ is a vector from a predetermined point (for example, the origin) to the position of the neuron u _IJ .

  Further, at time t + 1, the SOM processing unit 3 executes the same processing as described above.

  The time t + 1 indicates a time at which the data D2 is acquired by the feature vector acquiring unit 2 after the time t, and the acquired data D2 is input to the SOM processing unit 3.

  The SOM processing unit 3 repeatedly executes the above processing until time t = T.

The SOM processing unit 3 executes the above-described SOM processing, and acquires the reference vector vec_w _ij when the above-described SOM processing is completed, as the determined reference vector vec_w _ij _fixed. Then, SOM processor 3, the determined reference vector _{vec_w} ij _fixed, and outputs the mapper 7.

  In the following, for convenience of explanation, it is assumed that the input data vec_x, that is, the data D2 is composed of eight frequency component data, and the two-dimensional SOM has neurons arranged on 32 × 32 grid points. The case will be described as an example.

(1.2.2: Operation of template creation phase)
Next, the operation of the state determination device 1000 in the template creation phase will be described.

  In the template creation phase, the state determination device 1000 includes (1) template data in a walking state (walking state), (2) template data in a running state (running state), and (3) a standing state (standing). Is created).

  First, a case where template data in a walking state (walking state) is created will be described.

  In this case, an output from an acceleration sensor (not shown) when the person continues in the walking state (walking state) is input to the data input unit 1 as a signal Din.

  The data input unit 1 samples (A / D converts) the signal Din and obtains it as discrete data (digital data). Then, the data input unit 1 outputs the obtained discrete data (digital data) to the feature vector obtaining unit 2 as a signal D1.

  The feature vector acquisition unit 2 converts each of the digital data D1 by converting the discrete data D1 (digital data D1) acquired by the data input unit 1 into frequency domain data using, for example, Gabor wavelet transform. Acquire the intensity and phase of the frequency spectrum.

  Then, the feature vector acquisition unit 2 acquires, for example, data of a part of the band of the frequency spectrum intensity acquired by the above processing as the feature vector data D2.

  The feature vector data D2 acquired by the feature vector acquisition unit 2 is output from the feature vector acquisition unit 2 to the switch SEL1.

  A control unit (not shown) that controls each functional unit of the state determination device 1000 generates a selection signal sel1 for selecting the terminal 0 of the switch SEL1, and outputs the generated selection signal sel1 to the switch SEL1.

  The switch SEL1 outputs a signal D2 to the SOM processing unit 3 according to the selection signal sel1.

The SOM processing unit 3 performs SOM processing using the feature vector data D2 (signal D2) output from the switch SEL1. At this time, coupling weight vector (reference vector) vec_w _ij used in SOM process is a coupling weight vector (reference _{vector) vec_w} ij _fixed which was determined in the above learning phase.
SOM processor 3, the feature vector data D2 (signal D2), the connection weight vector (reference _{vector) vec_w} ij _fixed, executes the SOM process, acquires the SOM data D_som (t) of the current time t. Then, the SOM processing unit 3 outputs the acquired SOM data D_som (t) at the current time t to the SOM data storage unit 4.

  The SOM data storage unit 4 stores SOM output data D_som (t) output from the SOM processing unit 3.

  During the period from time t = 1 to t = N, the person is kept in the walking state (walking state), and the state determination device 1000 executes the above processing. Then, the SOM processing unit 3 stores the N pieces of SOM output data D_som (t) acquired during the period from time t = 1 to t = N in the SOM data storage unit 4.

  The template data generating unit 5 obtains N (N) acquired from the SOM data storage unit 4 during a period from time t = 1 to t = N (a period during which a person continues the walking state (walking state)). : Natural number) of SOM output data D_som (1) to D_som (N). Then, the template data generating unit 5 calculates the average value of the output values of the same coordinate position on the two-dimensional SOM (the output values of the same output node) using the SOM output data D_som (1) to D_som (N). I do. Then, template data in a walking state (walking state) is acquired by using the calculated average value as a value of each node (a node corresponding to an output node of the SOM).

That is, the feature vector data vec_x (t) at the time t is
vec_x (t) = (x ₁ (t), x ₂ (t),..., x _N1 (t))
N1 = 8
And complete the learning phase, and determine the determined connection weight vector f_vec_w _{ij
f_vec_w ij = (w ij_1, w} ij_2, ···, w ij_N1)
N1 = 8
And the output value of the neuron (output node) at the position (i, j) on the two-dimensional SOM at time t is D _ij (t),
N1 = 8
It is.

Note that w _{ij —} k is a connection weight between the output node at the coordinates (i, j) and the k-th input node.

The template data generation unit 5 obtains the value g _ij of the template data of the node corresponding to the neuron (output node) at the position of (i, j) on the two-dimensional SOM by the processing corresponding to the following equation.
The template data generation unit 5 acquires two-dimensional template data by assigning a value g _ij to the node at the coordinates (i, j) calculated by the above processing.

  FIG. 6 is a diagram illustrating an example of the template data.

The left diagram of FIG. 6 shows N (N: natural number) SOM output data obtained during a period from time t = 1 to t = N (a period during which a person continues a walking state (walking state)). This is template data Tmpl (1) (two-dimensional data composed of 32 × 32 data) acquired from D_som (1) to D_som (N) by the above processing. The value (pixel value) at the position of the coordinates (i, j) of the template data Tmpl (1) in the left diagram of FIG. 6 is the value g _ij calculated as described above.

In the template data shown in FIG. 6, the value g _ij is normalized to take a value from 0 to 1 (0 ≦ g _ij ≦ 1). In the template data shown in FIG. 6, the pixel of g _ij = 0 is a pixel of a gradation value “0” (pixel value “0” (pixel value corresponding to W0%)), and the pixel of g _ij = 1 Is shown as image data having pixels having a gradation value of “1” (pixel value “1” (pixel value corresponding to W100%)).

  As described above, the template data Tmpl (1) in the walking state (walking state) acquired by the template data generation unit 5 is stored in the template data storage unit 6.

  The process of creating the template data in the Running state (running state) (this template data is referred to as template data Tmpl (2)) is also performed in the same manner as described above.

  That is, during the period from time t = 1 to t = N, the acceleration sensor (not shown) when the person continues the Run state (running state) and the person continues the Running state (running state) ) Is input to the data input unit 1 as a signal Din. Then, the state determination device 1000 acquires the template data Tmpl (2) in the Running state (running state) by executing the same processing as described above. Then, the acquired template data Tmpl (2) in the Running state (running state) is stored in the template data storage unit 6.

  The center diagram in FIG. 6 shows N (N: natural number) SOMs acquired during a period from time t = 1 to t = N (a period in which a person continues the running state (running state)). This is template data Tmpl (2) (two-dimensional data composed of 32 × 32 data) acquired from the output data D_som (1) to D_som (N) by the above processing.

  The process of creating the template data in the standing state (standing state) (this template data is referred to as template data Tmpl (3)) is also performed in the same manner as described above.

  That is, during the period from time t = 1 to t = N, the acceleration sensor (not shown) when the person continues in the Standing state (standing state) and the person continues in the Standingg state (standing state) ) Is input to the data input unit 1 as a signal Din. Then, the state determination apparatus 1000 executes the same processing as described above to acquire template data Tmpl (3) in the standing state (standing state). Then, the acquired template data Tmpl (3) in the standing state (standing state) is stored in the template data storage unit 6.

  The right diagram in FIG. 6 shows N (N: natural number) SOMs acquired during a period from time t = 1 to t = N (a period during which the person continues the standing state (standing state)). This is template data Tmpl (3) (two-dimensional data composed of 32 × 32 data) acquired from the output data D_som (1) to D_som (N) by the above processing.

In the above description, the template data generation unit 5 uses the SOM output data D_som (1) to D_som (N) to average the output values of the same coordinate position on the two-dimensional SOM (the output values of the same output node). The case where template data is created by calculating a value has been described. However, the present invention is not limited to this method. For example, the template data generating unit 5 may create the template data by the following processes (“Process 1” to “Process 4”).
(1) Processing 1 (Threshold processing of output value)
Using the SOM output data D_som (1) to D_som (N), the template data generation unit 5 determines an output value (output value of the same output node) D _ij (t) at the same coordinate position on the two-dimensional SOM. , The average value of the output values D _ij (t) _satisfying D _ij (t) ≧ Th 1 is calculated, and the calculated average value is used as the value g _ij of the template data at the coordinates (i, j). .

  This processing will be described with reference to the flowchart of FIG.

  In step S1, the template data generation unit 5 sets a variable t representing a time, a counter value Cnt, and a total value Sum to “0”, respectively.

  In step S2, the template data generation unit 5 increments the variable representing the time by one.

Then, the template data generation unit 5 compares the output value (the output value of the same output node) D _ij (t) at the same coordinate position on the two-dimensional SOM with a predetermined threshold Th1, and D _ij (t) ≧ Th1 (In the case of “Yes” in step S3), the output value D _ij (t) is added to the total value Sum (step S4), and the counter value Cnt is incremented by 1 (step S5).

On the other hand, if D _ij (t) ≧ Th1 is not satisfied (“No” in step S3), the template data generating unit 5 advances the processing to step S6.

In step S6, if t ≧ N is not satisfied, the template data generating unit 5 returns the process to step S2, and if t ≧ N, the template data generating unit 5 substitutes the count value Cnt for the final count value N1. (Step S7), and
g _ij = Sum / N1
Then, the value g _ij of the template data at the coordinates (i, j) is obtained (step S8).

By performing the processing in this manner, the average value excluding the value whose SOM output value is smaller than the predetermined threshold value can be used as the value g _ij of the template data at the coordinates (i, j). Template data can be created.
(2) Processing 2 (threshold processing of variance value of output value)
The template data generation unit 5 uses the SOM output data D_som (1) to D_som (N) to distribute the output value (output value of the same output node) D _ij (t) at the same coordinate position on the two-dimensional SOM. The value v _ij is calculated by a process corresponding to the following equation.
Then, the template data generation unit 5 compares the variance value v _ij with a predetermined threshold Th2. Then, the template data generating unit 5 calculates an average value of the output values D _ij (t) of the SOM at the coordinates (i, j) satisfying v _ij ≤ Th2, and calculates the average value at the coordinates (i, j). Let the value of the template data be g _ij .

When v _ij ≦ Th2 is not satisfied, the template data generation unit 5 sets the value g _ij of the template data at the coordinates (i, j) to “0”.

By performing such processing, the average value calculated using only the value of the variance in the time direction of the SOM output value D _ij (t) that is small is set as the value g _ij of the template data at the coordinates (i, j). Therefore, more appropriate template data can be created. That is, by performing the processing as described above, the average value of the output values D _ij (t) of the SOM having little variation in the time direction can be used as the value g _ij of the template data of the coordinates (i, j). More appropriate template data can be created.
(3) Process 3 (combination process of process 1 and process 2)
The template data generation unit 5 may acquire the value g _ij of the template data at the coordinates (i, j) by a process combining the processes 1 and 2.

That is, the template data generation unit 5 compares the variance value v _ij with the predetermined threshold Th2. Then, the template data generation unit 5 determines that the output value D _ij (t) of the SOM at the coordinates (i, j) that satisfies v _ij ≦ Th2 and that the output value D _ij (t) is a predetermined threshold Th1 An average value of only the larger values is calculated, and the calculated average value is set as the value g _ij of the template data at the coordinates (i, j).

When v _ij ≦ Th2 is not satisfied, the template data generation unit 5 sets the value g _ij of the template data at the coordinates (i, j) to “0”.

As described above, even more appropriate template data can be created by the combination of the processing 1 and the processing 2.
(4) Processing 4 (processing using data within n1 × σ)
The template data generation unit 5 _calculates the variance value v _ij
σ _ij = sqrt (v _ij )
sqrt (x): The standard deviation σ _ij of the SOM output value D _ij (t) at the coordinates (i, j) is obtained by a process corresponding to a function for obtaining the square root of x.

When the template data generation unit 5 satisfies D _ij (t) −AveD _ij ≦ n1 × σ _ij (n1: a positive real number), the template data generation unit 5 calculates the SOM output value D _ij (t) of the coordinates (i, j). The average value is calculated, and the calculated average value is set as the value g _ij of the template data at the coordinates (i, j).

On the other hand, when the template data generating unit 5 does not satisfy abs (D _ij (t) −AveD _ij ) ≦ n1 × σ _ij (n1: a positive real number, abs (x) is a function for obtaining the absolute value of x) , The value g _ij of the template data at the coordinates (i, j) is set to “0”.

By performing such processing, template data can be created using data in a case where the output value D _ij (t) of the SOM is within a range of n1 × σ from the average value.

For example, when n1 = 3, the above processing is performed to create template data using data when the SOM output value D _ij (t) is within 3σ from the average value. can do. In the state determination device 1000, for example, when it is assumed that the output value D _ij (t) of the SOM is normally distributed, template data is created using data included in a range of 99.7% of the whole. , N1 = 3. As described above, the state determination device 1000 can adjust the range of the output value D _ij (t) of the SOM used for generating the template data by adjusting the value of n1.

  As described above, in the state determination device 1000, the template data is created.

Note that the template data may be normalized such that the sum of the values g _ij of the elements is “1”. By normalizing in this way, in the process of determining the degree of correlation between template data and SOM output data, no matter which template is used, a value indicating the degree of correlation (for example, the value of the inner product or the Euclidean distance) ) Can be constant (common), which facilitates the classification determination process.

(1.2.3: Operation of State Determination Processing Phase)
Next, the operation of the state determination processing phase of the state determination device 1000 will be described.

  FIG. 8 is a diagram (upper diagram) showing the waveform of the signal D1 acquired by the data input unit 1 (waveform in the walking state), and the feature vector data D2 acquired by the feature vector acquisition unit 2 from the signal D1 (lower left). FIG. 3) and SOM output data D_som (1) obtained from the feature vector data D2 by the mapping conversion unit 7.

  FIG. 9 is a diagram (upper diagram) showing the waveform of the signal D1 acquired by the data input unit 1 (the waveform in the Running state) (upper diagram), and the feature vector data D2 (lower left) acquired from the signal D1 by the feature vector acquisition unit 2. FIG. 5) and SOM output data D_som (2) obtained from the feature vector data D2 by the mapping conversion unit 7.

  An output from an acceleration sensor (not shown) for detecting the state of a person is input to the data input unit 1 as a signal Din. The acceleration sensor is mounted on, for example, a human body.

  The data input unit 1 samples (A / D converts) the signal Din and obtains it as discrete data (digital data). Then, the data input unit 1 outputs the obtained discrete data (digital data) to the feature vector obtaining unit 2 as a signal D1.

  For example, in the Walking state, a signal D1 as shown in the upper diagram of FIG. 8 is obtained. In the Running state, a signal D1 as shown in the upper diagram of FIG. 9 is obtained.

  The feature vector acquisition unit 2 acquires the feature vector data D2 from the discrete data D1 (digital data D1) acquired by the data input unit 1 in the same manner as the process in the learning phase.

  For example, in the Walking state, feature vector data D2 as shown in the lower left diagram of FIG. 8 is obtained. In the Running state, feature vector data D2 as shown in the lower left diagram of FIG. 9 is obtained.

  Then, the feature vector data D2 obtained by the feature vector obtaining unit 2 is output from the feature vector obtaining unit 2 to the switch SEL1.

  A control unit (not shown) that controls each functional unit of the state determination device 1000 generates a selection signal sel1 for selecting the terminal 1 of the switch SEL1, and outputs the generated selection signal sel1 to the switch SEL1.

  The switch SEL1 outputs the signal D2 to the mapping conversion unit 7 according to the selection signal sel1.

The mapping conversion unit 7 converts the feature vector data D2 (signal D2) by the coupling coefficient data w _ij _fixed between each node of the input layer and each neuron of the output layer determined by the SOM processing unit 3. Execute SOM processing. Then, the mapping conversion unit 7 outputs the SOM output data D_som acquired by the SOM processing to the matching processing unit 8.

  For example, in the Walking state, SOM output data as shown in the lower right diagram of FIG. 8 is obtained. In the Running state, SOM output data as shown in the lower right diagram of FIG. 9 is obtained.

  The template data acquisition unit 81 of the matching processing unit 8 reads from the template data storage unit 6 the template data Tmpl (1) in the Walking state, the template data Tmpl (2) in the Running state, and the template data Tmpl (3) in the Standing state. Is read.

  Then, the template data obtaining unit 81 outputs the read three template data Tmpl (1) to Tmpl (3) to the first inner product calculating unit 821 to the third inner product calculating unit 823, respectively.

  The first inner product calculation unit 821 executes a process of calculating an inner product using the SOM output data D_som and the template data Tmpl (1).

That is, the first inner product calculation unit 821 calculates the inner product by executing a process corresponding to the following equation.
Note that D _ij is the SOM output value of the coordinates (i, j) on the two-dimensional SOM, and g _ij is the value of the coordinates (i, j) of the template data Tmpl (1).

  The first inner product calculation unit 821 outputs a signal indicating the value of the inner product calculated as described above to the state determination unit 9 as a signal D_f (1).

  The second inner product calculation unit 822 performs a process of calculating an inner product using the SOM output data D_som and the template data Tmpl (2).

That is, the second inner product calculating unit 822 calculates the inner product by executing a process corresponding to the following equation.
Note that D _ij is the SOM output value of the coordinates (i, j) on the two-dimensional SOM, and g _ij is the value of the coordinates (i, j) of the template data Tmpl (2).

  The second inner product calculation unit 822 outputs a signal indicating the value of the inner product calculated as described above to the state determination unit 9 as a signal D_f (2).

  The third inner product calculation unit 823 performs a process of calculating an inner product using the SOM output data D_som and the template data Tmpl (3).

That is, the third inner product calculation unit 823 calculates the inner product by executing a process corresponding to the following equation.
Note that D _ij is the SOM output value of the coordinates (i, j) on the two-dimensional SOM, and g _ij is the value of the coordinates (i, j) of the template data Tmpl (3).

  The third inner product calculation unit 823 outputs a signal indicating the value of the inner product calculated as described above to the state determination unit 9 as a signal D_f (3).

  The matching processing unit 8 outputs the data obtained as described above to the state determination unit 9 as vector data vec_D_f (= (D_f (1), D_f (2), D_f (3))) of the fitness data.

  The state determination unit 9 receives the vector data vec_D_f of the goodness-of-fit data output from the matching processing unit 8. The state determination unit 9 determines the state at the current time t based on the input vector data vec_D_f.

  When the fitness data is acquired by the inner product as described above, the larger the fitness data (the value of the inner product), the higher the possibility that the state corresponds to the template used in the inner product processing.

  That is, D_f (1) has the largest value among the three pieces of fitness data (values of inner products) D_f (1), D_f (2), and D_f (3) acquired by the matching processing unit 8. In this case, it can be seen that the data input to the state determination device 1000 is close to the template data Tmpl (1) in the walking state, so that it is determined that the data input to the state determination device 1000 is the data in the walking state. Can be.

That is,
(1) When D_f (1) is the maximum value among the fitness data (values of inner products) D_f (1), D_f (2), and D_f (3), the state determination unit 9 A signal Dout indicating that the input data is data in the “walking state” is output to the outside.
(2) When D_f (2) is the maximum value among the fitness data (values of inner products) D_f (1), D_f (2), and D_f (3), the state determination unit 9 sends the state determination device 1000 A signal Dout indicating that the input data is data in the “Running state” is output to the outside.
(3) If D_f (3) is the maximum value among the fitness data (values of the inner product) D_f (1), D_f (2), and D_f (3), the state determination unit 9 A signal Dout indicating that the input data is data in the “Standing state” is output to the outside.

  For example, when the SOM data D_som (1) shown in FIG. 3 is input to the matching processing unit 8, D_f (1) becomes the maximum value. Therefore, the state determination unit 9 determines that the data is in the “walking state”. A signal Dout indicated is output to the outside.

  Then, when the SOM data D_som (2) shown in FIG. 3 is input to the matching processing unit 8, D_f (2) becomes the maximum value, and the state determination unit 9 determines that the data is in the “Running state”. A signal Dout indicated is output to the outside.

  As described above, the state determination device 1000 creates the template data indicating the predetermined state, calculates the value indicating the degree of correlation between the template data and the SOM output data (for example, the value of the inner product), and calculates the value. The state of the input data is determined based on the value indicating the degree of correlation.

  That is, in the state determination device 1000, even when a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the template indicating the predetermined state and the SOM output The degree of correlation (similarity) with data can be detected. Therefore, even if a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the state determination apparatus 1000 appropriately performs the pattern classification processing and the pattern determination processing. be able to.

<< First Modification >>
Next, a first modification of the first embodiment will be described.

  In this modification, the same parts as those in the first embodiment are denoted by the same reference numerals, and detailed description thereof will be omitted.

  FIG. 10 is a schematic configuration diagram of a state determination device 1000A according to a first modification of the first embodiment.

  FIG. 11 is a schematic configuration diagram of a matching processing unit 8A according to a first modification of the first embodiment.

  As shown in FIG. 10, a state determination device 1000A of the present modified example has a configuration in which the matching processing unit 8 is replaced with a matching processing unit 8A and the state determination unit 9 is replaced with the state determination unit 9A in the state determination device 1000 of the first embodiment. Has been replaced with

  The matching processing unit 8A of the present modification differs from the matching processing unit 8 of the first embodiment in that the first inner product calculating unit 821 to the M-th inner product calculating unit 82M are respectively replaced by a first distance calculating unit 831 to an M-th distance. It has a configuration in which the calculation unit 83M is replaced.

  The matching processing unit 8 of the first embodiment calculates the inner product to obtain the fitness data D_f. However, the matching processing unit 8A of the present modification calculates the Euclidean distance to convert the fitness data D_f. get.

That is, the first distance calculation unit 831 calculates the Euclidean distance between the SOM output data D_som and the template data Tmpl (1) by executing a process corresponding to the following equation.
Note that D _ij is the SOM output value of the coordinates (i, j) on the two-dimensional SOM, and g _ij is the value of the coordinates (i, j) of the template data Tmpl (1).

The second distance calculation unit 832 calculates the Euclidean distance between the SOM output data D_som and the template data Tmpl (2) by executing a process corresponding to the following equation.
Note that D _ij is the SOM output value of the coordinates (i, j) on the two-dimensional SOM, and g _ij is the value of the coordinates (i, j) of the template data Tmpl (2).

The third distance calculation unit 833 calculates a Euclidean distance between the SOM output data D_som and the template data Tmpl (3) by executing a process corresponding to the following equation.
Note that D _ij is the SOM output value of the coordinates (i, j) on the two-dimensional SOM, and g _ij is the value of the coordinates (i, j) of the template data Tmpl (3).

  The matching processing unit 8A outputs the acquired data as vector data vec_D_f (= (D_f (1), D_f (2), D_f (3))) of the fitness data to the state determination unit 9A.

  The state determination unit 9A inputs the vector data vec_D_f of the fitness data output from the matching processing unit 8A. The state determination unit 9A determines the state at the current time t based on the input vector data vec_D_f.

  When the fitness data is acquired from the Euclidean distance as described above, the smaller the fitness data (the value of the Euclidean distance), the higher the possibility of the state corresponding to the template used in the distance calculation processing.

  That is, D_f (1) has the smallest value among the three conformity data (inner product values) D_f (1), D_f (2), and D_f (3) acquired by the matching processing unit 8A. In this case, it can be seen that the data input to the state determination device 1000A is close to the walking state template data Tmpl (1), so that it is determined that the data input to the state determination device 1000A is the walking state data. Can be.

That is,
(1) If D_f (1) is the minimum value among the conformity data (values of the Euclidean distance) D_f (1), D_f (2), and D_f (3), the state determination unit 9A sets the state determination device 1000A. And outputs a signal Dout indicating that the data input to the DUT is data in the “walking state”.
(2) When D_f (2) is the minimum value among the conformity data (values of the Euclidean distance) D_f (1), D_f (2), and D_f (3), the state determination unit 9A sets the state determination device 1000A. And outputs a signal Dout indicating that the data being input to the is data in the “Running state”.
(3) If D_f (3) is the minimum value among the conformity data (values of the Euclidean distance) D_f (1), D_f (2), and D_f (3), the state determination unit 9A sets the state determination device 1000A. , Outputs a signal Dout indicating that the data input to the terminal is data in the “Standing state”.

  For example, when the SOM data D_som (1) shown in FIG. 11 is input to the matching processing unit 8, D_f (1) becomes the minimum value. Therefore, the state determination unit 9A determines that the data is in the “walking state”. A signal Dout indicated is output to the outside.

  When the SOM data D_som (2) shown in FIG. 11 is input to the matching processing unit 8, D_f (2) becomes the maximum value. Therefore, the state determination unit 9A determines that the data is in the “Running state”. A signal Dout indicated is output to the outside.

  As described above, the state determination device 1000A creates template data indicating a predetermined state, calculates a value indicating the degree of correlation between the template data and the SOM output data (for example, the value of the Euclidean distance), and calculates the value. The state of the input data is determined based on the value indicating the degree of correlation obtained.

  That is, in the state determination device 1000A, even when a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the template indicating the predetermined state and the SOM output The degree of correlation (similarity) with data can be detected. Therefore, even if a discontinuous area (for example, a divided area) exists in the map generated by the SOM, the state determination apparatus 1000A appropriately performs the pattern classification processing and the pattern determination processing. be able to.

[Other embodiments]
The state determination device may be configured by combining the above-described embodiment and modified examples.

  In the above-described embodiment and the modified example, the case where the SOM data storage unit 4 and the template data storage unit 6 are realized by separate memories has been described. However, the present invention is not limited to this. The unit 4 and the template data storage unit 6 may be realized by one memory.

  Further, in the state determination device described in the above embodiment, each block may be individually formed into one chip by a semiconductor device such as an LSI, or may be formed into one chip so as to include part or all.

  Note that, here, the LSI is used, but it may also be called an IC, a system LSI, a super LSI, or an ultra LSI depending on the degree of integration.

  Further, the method of circuit integration is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. An FPGA (Field Programmable Gate Array) that can be programmed after the LSI is manufactured, or a reconfigurable processor that can reconfigure connection and setting of circuit cells inside the LSI may be used.

  Further, a part or all of the processing of each functional block in each of the above embodiments may be realized by a program. Part or all of the processing of each functional block in each of the above embodiments is performed by a central processing unit (CPU) in a computer. Further, programs for performing the respective processes are stored in a storage device such as a hard disk or a ROM, and are executed by being read from the ROM or from the RAM.

  Further, each process of the above embodiment may be realized by hardware, or may be realized by software (including a case where it is realized together with an OS (operating system), middleware, or a predetermined library). Further, it may be realized by mixed processing of software and hardware.

  For example, when each functional unit of the above-described embodiment (including modified examples) is realized by software, the hardware configuration (for example, a CPU, a ROM, a RAM, an input unit, an output unit, and the like) illustrated in FIG. Each functional unit may be realized by software processing using a connected hardware configuration).

  The execution order of the processing method in the above embodiment is not necessarily limited to the description of the above embodiment, and the execution order can be changed without departing from the gist of the invention.

  A computer program that causes a computer to execute the above-described method and a computer-readable recording medium that records the program are included in the scope of the present invention. Here, examples of the computer-readable recording medium include a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, large-capacity DVD, next-generation DVD, and semiconductor memory. .

  The computer program is not limited to one recorded on the recording medium, and may be transmitted via a telecommunication line, a wireless or wired communication line, a network represented by the Internet, or the like.

  In addition, the word “part” may be a concept including “circuitry”. The circuit may be realized in whole or in part by hardware, software, or a mixture of hardware and software.

  The specific configuration of the present invention is not limited to the above embodiment, and various changes and modifications can be made without departing from the spirit of the invention.

1000, 1000A State determination device 1 Data input unit 2 Feature vector acquisition unit 3 SOM processing unit 5 Template data generation unit 7 Mapping conversion unit 8 Matching processing unit 9 State determination unit

Claims (12)

A feature vector acquisition unit that acquires feature vector data from measurement data obtained by measuring an event in an unknown state,
A mapping conversion unit that obtains SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data;
A matching processing unit that obtains fitness data indicating a degree of correlation between the template data indicating a predetermined state and the SOM output data obtained by the mapping conversion unit;
A state determination unit that determines a state indicated by the measurement data based on the fitness data acquired by the matching processing unit,
Equipped with a,
The matching processing unit,
An inner product of each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data is calculated, and the calculated inner product value is converted into M (M: natural number) templates. Acquisition as fitness data D_f (1) to D_f (M) indicating the degree of correlation between each of the data Tmpl (1) to Tmpl (M) and the SOM output data,
The state determination unit,
Among the fitness data D_f (1) to D_f (M), the fitness data having the maximum value is the fitness data indicating the highest correlation, and the fitness indicating the highest correlation is used as the fitness data indicating the highest correlation. A state determination device that specifies the template data corresponding to data and determines that a state corresponding to the specified template data is a state indicated by the measurement data . A feature vector acquisition unit that acquires feature vector data from measurement data obtained by measuring an event in an unknown state,
A mapping conversion unit that obtains SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data;
A matching processing unit that obtains fitness data indicating a degree of correlation between the template data indicating a predetermined state and the SOM output data obtained by the mapping conversion unit;
A state determination unit that determines a state indicated by the measurement data based on the fitness data acquired by the matching processing unit,
Equipped with a,
The matching processing unit,
The Euclidean distance between each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data is calculated, and the calculated Euclidean distance value is set to M (M: natural number) Are obtained as fitness data D_f (1) to D_f (M) indicating the degree of correlation between each of the template data Tmpl (1) to Tmpl (M) and the SOM output data,
The state determination unit,
Among the fitness data D_f (1) to D_f (M), the fitness data having the smallest value is the fitness data indicating the highest correlation, and the fitness indicating the highest correlation is used as the fitness data indicating the highest correlation. A state determination device that specifies the template data corresponding to data and determines that a state corresponding to the specified template data is a state indicated by the measurement data . The template data is data composed of two-dimensional elements,
The sum of the values of the respective element data of the template data is “1”.
State determining apparatus according to claim 1 or 2. The template data is data composed of two-dimensional elements,
The template data is data generated based on N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) in which the same state continues,
When the template data is data composed of m × n (m, n: natural numbers),
Let g _{ij be} the element data of the two-dimensional coordinates (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) of the template data,
Assuming that the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is D _ij (t),
The template data is

Have been the value g _ij obtained by two-dimensional coordinates of the template (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) generated by the element data Data
State determining apparatus according to any one of claims 1 to 3. The template data is data composed of two-dimensional elements,
The template data is data generated based on N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) in which the same state continues,
When the template data is data composed of m × n (m, n: natural numbers),
Let g _{ij be} the element data of the two-dimensional coordinates (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) of the template data,
Assuming that the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is D _ij (t),
The template data compares the element data D _ij (t) with a predetermined threshold Th1, calculates an average value of the element data D _ij (t) that satisfies D _ij (t) ≧ Th1, and coordinates the calculated average value with the coordinates. The value g _ij of the template data of (i, j) is the generated data.
State determining apparatus according to any one of claims 1 to 3. The template data is data composed of two-dimensional elements,
The template data is data generated based on N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) in which the same state continues,
When the template data is data composed of m × n (m, n: natural numbers),
Let g _{ij be} the element data of the two-dimensional coordinates (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) of the template data,
Assuming that the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is D _ij (t),
The template data is

The variance value vij of the element data at the coordinates (i, j) obtained by the above is compared with a predetermined threshold Th2, and the value D _{ij of the} SOM output data at the coordinates (i, j) satisfying v _ij ≤ Th2 is satisfied. t) is calculated, and the calculated average value is used as the value g _ij of the template data at the coordinates (i, j) to generate data.
State determining apparatus according to any one of claims 1 to 3. The template data is data composed of two-dimensional elements,
The template data is data generated based on N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) in which the same state continues,
When the template data is data composed of m × n (m, n: natural numbers),
Let g _{ij be} the element data of the two-dimensional coordinates (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) of the template data,
Assuming that the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is D _ij (t),
The template data is


The variance value v _ij of the element data at the coordinates (i, j) obtained by the above is compared with a predetermined threshold Th2, and the value D _{ij of the} SOM output data at the coordinates (i, j) satisfying v _ij ≦ Th2 (T), and calculate the average value of only the SOM output data whose value D _ij (t) is larger than a predetermined threshold value Th3, and use the calculated average value as a template of coordinates (i, j). By generating the data value g _ij , the generated data
State determining apparatus according to any one of claims 1 to 3. The template data is data composed of two-dimensional elements,
The template data is data generated based on N pieces of SOM output data acquired during a period from time t = 1 to time t = N (N: natural number) in which the same state continues,
When the template data is data composed of m × n (m, n: natural numbers),
The element data of the two-dimensional coordinates (i, j) (i, j: natural number, 1 ≦ i ≦ m, 1 ≦ j ≦ n) of the template data is defined as gij,
Assuming that the element data of the two-dimensional coordinates (i, j) of the SOM output data at time t is D _ij (t),
The template data is

σ _ij = sqrt (v _ij )
sqrt (x): When the standard deviation of the element data at the coordinates (i, j) obtained by the function for obtaining the square root of x is σ _ij ,
When abs (D _ij (t) −AveD _ij ) ≦ n1 × σ _ij (n1: a positive real number, abs (x) is a function for obtaining the absolute value of x), the SOM of the coordinates (i, j) is satisfied. The average value of the output values D _ij (t) is calculated, and the calculated average value is used as the value g _ij of the template data at the coordinates (i, j).
If abs (D _ij (t) −AveD _ij ) ≦ n1 × σ _ij is not satisfied, the value g _ij of the template data at the coordinates (i, j) is set to “0”, which is data generated.
State determining apparatus according to any one of claims 1 to 3. A feature vector acquisition step of acquiring feature vector data from measurement data obtained by measuring an event in an unknown state,
A mapping conversion step of obtaining the SOM output data by projecting the feature vector data onto a space having a dimension different from the dimension of the feature vector data;
A matching processing step of acquiring template data indicating a predetermined state, and fitness data indicating a degree of correlation between the SOM output data acquired in the mapping conversion step,
A state determination step of determining a state indicated by the measurement data based on the fitness data acquired by the matching processing step;
Equipped with a,
The matching processing step includes:
An inner product of each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data is calculated, and the calculated inner product value is converted into M (M: natural number) templates. Acquisition as fitness data D_f (1) to D_f (M) indicating the degree of correlation between each of the data Tmpl (1) to Tmpl (M) and the SOM output data,
The state determination step includes:
Among the fitness data D_f (1) to D_f (M), the fitness data having the maximum value is the fitness data indicating the highest correlation, and the fitness indicating the highest correlation is used as the fitness data indicating the highest correlation. A program for causing a computer to execute a state determination method of specifying the template data corresponding to data and determining a state corresponding to the specified template data as a state indicated by the measurement data . A feature vector acquisition step of acquiring feature vector data from measurement data obtained by measuring an event in an unknown state,
A mapping conversion step of obtaining the SOM output data by projecting the feature vector data onto a space having a dimension different from the dimension of the feature vector data;
A matching processing step of acquiring template data indicating a predetermined state, and fitness data indicating a degree of correlation between the SOM output data acquired in the mapping conversion step,
A state determination step of determining a state indicated by the measurement data based on the fitness data acquired in the matching processing step;
Equipped with a,
The matching processing step includes:
The Euclidean distance between each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data is calculated, and the calculated Euclidean distance value is set to M (M: natural number) Are obtained as fitness data D_f (1) to D_f (M) indicating the degree of correlation between each of the template data Tmpl (1) to Tmpl (M) and the SOM output data,
The state determination step includes:
Among the fitness data D_f (1) to D_f (M), the fitness data having the smallest value is the fitness data indicating the highest correlation, and the fitness indicating the highest correlation is used as the fitness data indicating the highest correlation. A program for causing a computer to execute a state determination method of specifying the template data corresponding to data and determining a state corresponding to the specified template data as a state indicated by the measurement data . A feature vector acquisition unit that acquires feature vector data from measurement data obtained by measuring an event in an unknown state,
A mapping conversion unit that obtains SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data;
A matching processing unit that obtains fitness data indicating a degree of correlation between the template data indicating a predetermined state and the SOM output data obtained by the mapping conversion unit;
A state determination unit that determines a state indicated by the measurement data based on the fitness data acquired by the matching processing unit,
Equipped with a,
The matching processing unit,
An inner product of each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data is calculated, and the calculated inner product value is converted into M (M: natural number) templates. Acquisition as fitness data D_f (1) to D_f (M) indicating the degree of correlation between each of the data Tmpl (1) to Tmpl (M) and the SOM output data,
The state determination unit,
Among the fitness data D_f (1) to D_f (M), the fitness data having the maximum value is the fitness data indicating the highest correlation, and the fitness indicating the highest correlation is used as the fitness data indicating the highest correlation. An integrated circuit that specifies the template data corresponding to data and determines that a state corresponding to the specified template data is a state indicated by the measurement data . A feature vector acquisition unit that acquires feature vector data from measurement data obtained by measuring an event in an unknown state,
A mapping conversion unit that obtains SOM output data by projecting the feature vector data into a space having a dimension different from the dimension of the feature vector data;
A matching processing unit that obtains fitness data indicating a degree of correlation between the template data indicating a predetermined state and the SOM output data obtained by the mapping conversion unit;
A state determination unit that determines a state indicated by the measurement data based on the fitness data acquired by the matching processing unit,
Equipped with a,
The matching processing unit,
The Euclidean distance between each of the M (M: natural number) template data Tmpl (1) to Tmpl (M) and the SOM output data is calculated, and the calculated Euclidean distance value is M (M: natural number) Are obtained as fitness data D_f (1) to D_f (M) indicating the degree of correlation between each of the template data Tmpl (1) to Tmpl (M) and the SOM output data,
The state determination unit,
Among the fitness data D_f (1) to D_f (M), the fitness data having the smallest value is the fitness data indicating the highest correlation, and the fitness indicating the highest correlation is used as the fitness data indicating the highest correlation. An integrated circuit that specifies the template data corresponding to data and determines that a state corresponding to the specified template data is a state indicated by the measurement data .