JPH02244097A - Time series signal corresponding device - Google Patents

Abstract

PURPOSE:To allow the exact correspondence of the significant components pertinent to the same cause by corresponding the significant components pertinent to the same cause by using a dictionary in which the information on the significant components relating to hypothetic causes are stored with respect to plural time series signals. CONSTITUTION:A characteristic quantity extracting means 20 detects the significant components of the respective time series signals inputted in input sections 11 to 1N, extracts characteristic quantities outputs to a similarity information calculating means 30, and outputs the information on the significant components respectively to a coupling weight information calculating means 30. The characteristic dictionary 40 outputs the hypothetic causes to generate the significant components in the respective time series signals and the information on the characteristics to a similarity information calculating means 30. The similarity information calculating means 30 calculates the similarity information of the corresponding combination taking out the significant components by one piece each from the respective time series signals by using the information stored in the characteristic dictionary 40 and the characteristic quantities. The coupling weight information calculating means 30 calculates the information on the coupling weights and determines the corresponding combination of the significant components pertinent to the same cause in the classifying means 50. The exact correspondence of the significant components pertinent to the same cause is executed in this way.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to a time-series signal mapping device for detecting corresponding components in a plurality of time-series signals.

[Prior art] As an algorithm for associating multiple time-series signals, regardless of whether they are analog signals or digital signals, there is a method based on dynamic programming (co-authored by Yasuin and Chubu, [Computer Speech Processing J, published by Akiba Publishing, 1988]
January 6, ppI68-170). This method takes into account the corresponding parts (hereinafter, the corresponding subelements are referred to as significant components) even if there is missing data, insertion of extra data, or data errors in the input series, and finds the optimal correspondence. This is what we intend to do.

[Problem to be Solved by the Invention] According to this method, even if the above-mentioned abnormal situation occurs, it is possible to accurately associate the significant components, but the premise is that the order of the significant components is not changed. , it is not possible to associate multiple time-series signals with different orders of significant components.

For example, for multiple light receiving units installed at different locations that receive light from multiple light sources that turn on and off independently,
Since the distances between each light source and each light receiving section are different, the light beam from the light source that is first received by a certain light receiving section may be received second or third by another light receiving section. That is, the order of significant components of the time-series signals output from each light receiving section may be different.

There are devices that are required to match the significant components related to the same cause even for multiple time-series signals with different orders of significant components, but in the past, it is necessary to match time-series signals with different orders of significant components. No equipment is proposed.

The present invention has been made in consideration of the above points, and is a time-series signal mapping method that can accurately correlate significant components even for a plurality of time-series signals in which the order of significant components related to the same cause is different. is intended to provide a device.

[Means for Solving the Problem] In order to solve the problem, the present invention provides a plurality of time-series signals having a plurality of significant components on the time axis, and different time-series signals have significant components due to the same cause. For multiple time-series signals whose components may appear in different orders,
The apparatus for associating significant components related to the same cause with time-series signals was constructed using the following means.

That is, a feature extracting means for extracting a feature for a significant component, an assumed cause that causes a significant component to occur in each time-series signal, and characteristic information of the significant component of each time-series signal caused by the assumed cause, and a characteristic dictionary storing a plurality of hypothetical causes. In addition, regarding multiple significant components included in each time series signal within the analysis time width,
For corresponding combinations from which one significant component is extracted from each time-series signal, similarity information of multiple significant components of the corresponding combinations is obtained using information stored in the characteristic dictionary and feature quantities extracted by the feature quantity extraction means. and a connection weight information calculation means for obtaining connection weight information when a certain set of corresponding combinations and another set of corresponding combinations are connected. The apparatus further includes a classification means for determining corresponding combinations of significant components related to the same cause based on similarity information and connection weight information.

[Operation] The feature extracting means detects the significant component of each time-series signal, extracts the feature for the detected significant component, outputs the extracted feature to the similar information calculating means, and detects the significant component. Information indicating this is output to the connection weight information calculation means. The characteristic dictionary stores assumed causes that cause significant components to occur in each time-series signal, and characteristic information of the significant components of each time-series signal that occur due to the assumed causes, and these stored information are used by the similar information calculation means. Output to.

The similar information calculation means calculates, for a plurality of significant components included in each time-series signal within the analysis time width, a plurality of corresponding combinations from which one significant component is extracted from each time-series signal. Similar information of significant components,
The similarity information is calculated using the information stored in the characteristic dictionary and the feature extracted by the feature extraction means, and the calculated similarity information is output to the classification means. The connection weight information calculation means calculates connection weight information when one set of corresponding combinations is connected with another set of corresponding combinations, and outputs the calculated connection weight information to the classification means.

The classification means determines corresponding combinations of significant components related to the same cause based on the similarity information and the connection weight information.

[Example] Hereinafter, an example of the present invention will be described in detail with reference to the drawings.

Here, FIG. 1 is a block diagram showing the overall configuration of this embodiment, and FIG. 2 is a signal waveform diagram for explaining the feature amount of the significant component of the input time-series signal.

real blood war week death First, the configuration of the embodiment will be explained.

In FIG. 1, a plurality of input units 11 to IN each input a time series signal, and each time series signal includes significant components caused by the same cause. Therefore,
If there are multiple causes, each time-series signal will include multiple significant components.

Here, the order of causes for significant components in one time-series signal does not necessarily match the order of causes for significant components in other time-series signals, and each time-series signal related to the same cause is processed by the following processing means. Classify by associating significant components. The time-series signals inputted by these input units 11 to IN are provided to a feature extraction unit 20.

For each time series signal, the feature extraction unit 20
It extracts special quantities such as time, power, frequency, etc., and provides the extracted special quantities to the coefficient calculating section 30.

The coefficient calculation unit 30 is given characteristic information from a characteristic dictionary 40 . In the characteristic dictionary 40, when it is assumed that a certain cause that causes a significant component occurs, each input section 11 to
Characteristic information about significant components appearing in time-series signals from is stored for a plurality of hypothetical causes, and is provided to the coefficient calculation unit 30. The coefficient calculation section 30 uses the feature amount information about each significant component given from the feature amount extraction section 20 and the characteristic information given from the characteristic dictionary 40 to calculate each time series executed by the classification section 50, which will be described later. A combination weighting coefficient, a corresponding fixed value, etc., which are used in a calculation for associating significant components related to the same cause of signals, are calculated.

The classification unit 50 classifies significant components that are related to the same cause in each time-series signal by repeatedly performing calculations as described later using the given connection weighting coefficients, corresponding fixed values, and the like.

Operation of this embodiment The operation of this embodiment having the above configuration will be described below. In addition, in order to simplify the explanation, the input section 11 to IN is 2.
Individual.

The time-series signals inputted from each of the input units 11 and 12 are respectively given to the feature amount extraction unit 20, and the feature amount extraction unit 20 extracts the feature amount for each significant component.

Extraction of this feature amount will be explained with reference to FIG.

The feature extraction unit 20 first calculates the power change shown in FIG. 2(B) from each input time series signal shown in FIG. 2(A), and the calculated power is set to a predetermined threshold value TH. The interval t3 to 1o that exceeds the above is defined as a significant component interval. Then, the significant component time ta at which the power is maximum within this interval t, ~t8, and the time length TL (=t, -
t8), the maximum power value P, and this interval t, ~t8
Fourier transform output (frequency characteristics) of the input time series signal
(ω) is extracted as a feature quantity. The feature amount information extracted in this way is given to the coefficient calculating section 30.

Characteristic information is also given to the coefficient calculating section 30 from the characteristic dictionary 40 .

The characteristic information stored in the characteristic dictionary 40 given to the coefficient calculation unit 30 includes a plurality of time characteristic functions (11, 12
) and a plurality of power characteristic functions Gr(pl, P2).

Here, the time characteristic function Dr (tl, t2) is an ideal value for the significant component appearing in the time series signal from each input section 11.12, assuming that a certain cause r that causes the significant component has occurred. Significant component times t1r, t2r
It is a function that represents the error between the significant component times t1 and t2 that were actually obtained, and it is close to 0 when the actual cause content and the assumed cause content are similar; The larger the positive value is, the more the content of the cause differs. For example, ideal significant component times t16, t2r
Significant component time t actually obtained for the theoretical formula for
1, a function that outputs the error when t2 is approximated by least squares can be applied. Power characteristic function G, (Pl,
P,2) is also a similar function for the maximum powers P1, P2.

Various coefficients are calculated from such time characteristic functions (tl, t2 >, power characteristic functions G, (Pl, P2), and the feature information extracted by the feature extraction unit 20. Before explaining the meanings and calculation methods of these coefficients, the classification process performed by the classification unit 50 using these coefficients will be described.

For two time-series signals having multiple significant components corresponding to multiple causes, classification processing that associates significant components of the same cause is performed by dividing each time-series signal into fixed analysis time widths. Ru.

Now, let the numbers of significant components included in each time-series signal within a certain analysis time width be ml and ml, respectively, and let the number of significant components m1 be m, and the number of small ml be m, then the classification unit 50 calculates the following:
A process of classifying and determining a maximum of m sets of significant components due to the same cause is performed.

The classification process by the classification unit 50 is performed by the following recursive iterative calculation (n indicates the n-th iterative calculation).

Here, x, p are 1.2,...ml, and y.

q is 1.2,...ml.

Further, Vx in equation (1) indicates the corresponding state value when the X-th significant component of the first time-series signal and the y-th significant component of the second time-series signal are associated. , this corresponding state value V xy is a value in the range of O to 1 in this embodiment, and the closer it is to 1, the more the This indicates that it is highly likely that the y-th significant component is related to the same cause.

W in equation (1) is the first time series signal Xy,
Correspondence state value Vi between the X-th significant component of the fourth signal and the y-th significant component of the second time-series signal, and the first significant component of the first time-series signal and the second time-series signal This is a weighting coefficient for the connection between the ninth significant component of the signal and the corresponding state value Vpq.

When the X-th significant component of the first time-series signal is associated with the y-th significant component of the second time-series signal, this association also affects the association with other significant components. In addition, it is also influenced by the correspondence of other significant components, and only the correspondence between the X-th significant component of the first time-series signal and the y-th significant component of the second time-series signal is determined. It is not enough to simply consider this, but the correspondence between other significant components must also be considered with a certain weight.

Therefore, the product-sum operation ΣΣWX'/, I)Q in (1)
・As in V4, the correspondence between other significant components is given a certain weight, and the X-th significant component of the first time-series signal and the y-th significant component of the second time-series signal are This is reflected in the corresponding state value Vxy.

In addition, in this way, the correspondence state value ■ between the Xth significant component of the first time series signal and the yth significant component of the second time series signal reflects other correspondences y of the significant components. However, it is also necessary to reflect information on these correlated significant components without considering information on other significant components. Ix in formula (1) is
This is a corresponding eigenvalue determined from information on the X-th significant component of the first time-series signal and information on the y-th significant component of the second time-series signal.

Furthermore, the function f (z) in equation (1) is a threshold numerical aperture that has a characteristic that the output value is saturated with respect to the input value. As this threshold numerical aperture f (z), for example, a sigmoid function shown in the following equation (2) can be applied.

Note that shi is a predetermined constant.

Therefore, when the classification unit 50 attempts to repeatedly calculate the recursive formula shown in equation (1), the connection weight coefficient WXV, l)
Q and the corresponding eigenvalue I, are required. The coefficient calculation unit 30 calculates these values and provides them to the classification unit 50.

The connection weighting coefficient W is determined by the following formula. Here, A, B, and C are predetermined constants, δij is 1 when i and j are equal, and δij is 1 when i and j are equal.
The value is 0 if the values are different.

However, if X and p are equal and y and 9 are equal, that is, the connection weighting coefficient WX for the same corresponding state values
V, I) Q is 0.

As is clear from equation (3), the connection weighting coefficient is
The connection weighting coefficient W becomes -C, 2 pieces XV, i)
Corresponding state value V of Q, ,■9. (x=p, y≠q), the connection weighting coefficient W is -A
-C, and the two corresponding state values V =
In q), the connection weighting coefficient W is -Bxy, pq -C, and this connection weighting coefficient W is
is changing the value of This formula (3) is (1)XV,
It is determined from the contents of the pQ formula, and as a result, the two corresponding state values that map one significant component of one time series signal to different significant components of the other time series signal. The connection weighting coefficient has a different value from other connection weighting coefficients.

In addition, the above-mentioned corresponding eigenvalues ■8 used for classifying the corresponding significant components. is obtained according to the following formula % formula % (4). Here, D is a predetermined constant, and d, X, and y represent the degree of dissimilarity between the Xth significant component of the first time series signal and the yth significant component of the second time series signal. This is the value shown. As shown in equation (4), the correspondence fixed value xV is affected by the number m of significant components within the analysis period width, and the correspondence between each significant component that is about to be reduced by one dimension. This takes into consideration the fact that the correspondence is influenced by the degree of similarity (dissimilarity) between the two.

The dissimilarity dX, which is the element that determines this fixed correspondence value I,
y is the time characteristic function for the hypothetical cause r (r is 1.2, . . . S) stored in the characteristic dictionary 40, and (tl
, t2) and the power characteristic function G.

(Pl, P2) and the time t for the Xth significant component of the first time series signal given from the feature scene extraction unit 20
ax, significant component time length TLX, maximum power value P frequency characteristic Value P01, frequency characteristic χ
It is obtained by calculation using the following formula (%) using V (ω). However, w1 to W5 in this equation (5) are predetermined constants.

The first term in Equation (5) is the assumed maximum This shows the amount of error regarding time and power for causes with similar content. The second term indicates the difference in frequency characteristics between the X-th significant component of the first time-series signal and the y-th significant component of the second time-series signal. The third term indicates the difference in time length between the X-th significant component of the first time-series signal and the y-th significant component of the second time-series signal.

The connection weight coefficient W calculated by the coefficient calculation unit 30
and corresponding eigenvalue ■8. is given to the classification unit 50 in an xyl concave manner as described above. The classification unit 50 uses these to repeat the calculation of equation (1), which is not described above.

Here, the initial value V of the corresponding state value used for the first calculation
, (1) uses any value in the range of 0 to 1 (which may be a random value). In addition, the iterative calculation of the formula (1), which is not shown in formula (1), may be performed a fixed number of times,
Further, the process may be performed until the change sound L7 of the corresponding state value becomes smaller than the threshold value, and further, the process may be performed until the number classified as corresponding becomes a predetermined ratio of the maximum possible number of corresponding items m.

Through such repeated calculations, the correspondence status value between significant components related to the same cause approaches 1, and the correspondence status value between significant components related to different causes approaches O. Classification section 50
At the end of the iterative calculation, the two significant components associated with the corresponding state value having a value closer to 1 than a predetermined threshold are output as corresponding components.

Therefore, according to the above-described embodiment, it is possible to classify a plurality of time-series signals in which significant components related to the same cause may appear in different orders by accurately correlating the significant components related to the same cause.

In this case, since the characteristic dictionary 40 is used, the device can be made versatile by changing the contents of the dictionary. Also, since the characteristic dictionary 40 is used,
It is extremely unlikely that noise components mixed into some time series signals will be erroneously associated with significant components of other time series signals.

In addition, in the explanation of the operation in the above-mentioned embodiment, an example in which there are two input units is shown, but the same can be applied to a case in which there are three or more input units.

For reference, the connection weight coefficient Wxy in three cases
The calculation formulas for z and pqr are shown below.

W =-A1 ・δxp(1-δyq'δzr
) xyz, pqr A2 ・δ7. (1-δ8.・δ7.)-A3゛δl''
(1-δxp°δyq)A4...
(6) In the above embodiment, the degree of similarity (dissimilarity) is set to '4. Although we have shown calculations using all four elements: the time when the significant component occurs, the significant component time length, the maximum power value, and the frequency characteristics, it is also possible to calculate using at least some elements that include time information. Also good. for example,
In the case of a time-series signal whose significant component is a rectangular pulse with a fixed amplitude, there is no need to consider the maximum power value or frequency characteristics.

Further, in the above-described embodiment, a sigmoid function is used as the threshold function, but other functions may be used. In other words, it may be a function that monotonically increases as the input increases, and saturates to a large value when the input is above a certain value, and saturates to a small value when the input is below a certain value.

[Effects of the Invention] As described above, according to the present invention, among a plurality of time-series signals having a plurality of significant components on the time axis, the order in which the significant components due to the same cause appear between different time-series signals is For multiple time-series signals that may differ, significant components related to the same cause are matched using a dictionary that stores significant component information related to hypothesized causes, so it is not possible to accurately match significant components related to the same cause. can.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the overall configuration of an embodiment of the time-series signal mapping apparatus according to the present invention, and FIG. 2 is a signal waveform diagram for explaining the specific quantity of significant components of input time-series signals. . 11-1n...input section, 20...feature amount extraction section, 3
0... Coefficient calculation unit, 40... Characteristic dictionary, 50...
Classification department. 11~IN, signal waveform diagram Fig. 2 for explaining input section feature quantities

Claims (3)

[Claims] (1) For multiple time-series signals that have multiple significant components on the time axis, the order in which the significant components appear due to the same cause may differ between different time-series signals. A time-series signal associating device for associating significant components with each other, comprising: a feature extracting means for extracting a feature for the significant component; a hypothetical cause for causing the significant component in each time-series signal;
A characteristic dictionary that stores characteristic information of significant components of each time-series signal caused by assumed causes for multiple hypothesized causes, and a characteristic dictionary that stores characteristic information of significant components of each time-series signal caused by assumed causes, and information about multiple significant components included in each time-series signal within the analysis time width. , for corresponding combinations from which one significant component is extracted from each time-series signal, similarity information of a plurality of significant components of the corresponding combinations is extracted from information stored in the characteristic dictionary and feature quantities extracted by the feature quantity extraction means. a means for calculating similarity information obtained using the following methods; a means for calculating connection weight information for obtaining connection weight information when one set of corresponding combinations is connected with another set of corresponding combinations; and a classification means for determining corresponding combinations of significant components related to the same cause. (2) The above classification means is the recursive formula of the following formula ▲ There are mathematical formulas, chemical formulas, tables, etc. ▼ (Here, a and b indicate one out of c corresponding combinations, and n indicates the number of repeated operations. where V_a(n) is the corresponding state value at the number n of repeated operations for the corresponding combination a, and when each significant component of the corresponding combination a is related to the same cause, it approaches a certain value as the repeated operations progress, W_a_, _b indicate the combination weight information of the corresponding combinations a and b, I_a indicates the similarity information of each significant component of the corresponding combination a, and V_b (
n-1) indicates the corresponding state value at the number of iterative operations n-1 for the corresponding combination b, and f() is a threshold function with saturated input-output characteristics. , it is determined whether the corresponding combination a is a significant component related to the same cause based on the corresponding state value V_a(n) at the end of the iterative operation. Series signal matching device. (3) The characteristic dictionary stores at least characteristic information regarding time information from the occurrence of a certain cause until the occurrence of a significant component corresponding to the cause in each time-series signal, and the similar information calculation means 2. The time-series signal correlating device according to claim 1, wherein the similarity information is calculated using at least characteristic information regarding the time information.