JP2015212635A - Threshold waveform creation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a threshold waveform creation device capable of creating a threshold waveform considering genuine variation, in which a variation in a time axis direction of an observation waveform does not appear as a variation in an observation value axis direction, by obtaining a genuine average waveform in which characteristics of the waveform is not lost due to an effect of smoothing by averaging of the observation waveform.SOLUTION: The threshold waveform creation device includes: a feature extraction section 13 that extracts feature points from an observation waveform; a correlation evaluation part 14 that evaluates the feature point extracted by the feature extraction section 13; a statistical processing section 15 that performs a statistical processing on an observation waveform in a time axis direction and in an observation value axis direction based on the feature point; and a threshold waveform creation section 16 that creates a threshold waveform which is used for determining the normality of the observation waveform based on the values after subjected to statistical processing in the time axis direction and the observation value axis direction.

Description

  The present invention relates to a threshold waveform creating apparatus that creates a threshold waveform used to determine the normality of an observed waveform.

  Conventionally, in order to prevent a product containing a defect from being shipped to the market, an inspection for determining the quality of a manufactured product has been performed. As one of the inspection methods, a test signal is given to a manufactured product, the behavior of the product at that time is measured through a sensor or the like, and a measurement waveform acquired as a time series is a threshold or threshold waveform that is a predetermined reference. There is known a method for judging the quality of a product as compared with the above.

  For example, a determination method using a threshold waveform is based on whether the acquired waveform deviates from an upper limit waveform or a lower limit waveform of a threshold waveform set in advance. Since this determination method only needs to set the threshold waveform, it can be applied in various fields. This is especially beneficial when multi-variety variable-volume production is carried out.

  Patent Documents 1 and 2 describe a method for creating a threshold waveform. In the method disclosed in Patent Document 1, an average waveform calculated from a plurality of normal waveforms is added with a standard deviation of the plurality of normal waveforms to obtain an upper limit waveform, and subtracted to obtain a lower limit waveform. On the other hand, the method disclosed in Patent Document 2 does not evaluate the average or standard deviation of a plurality of normal waveforms, but defines a space surrounded by a plurality of normal waveforms as a space to be determined as good. An upper limit waveform and a lower limit waveform are created.

JP 2002-341909 A JP 2004-239879 A

  However, in the method proposed in Patent Document 1, an average waveform obtained by averaging a plurality of acquired normal waveforms is not necessarily a true average normal waveform. In order to be truly average, it is necessary to acquire and average a certain number of normal waveforms, and it is not sufficient that the acquired number is about 1-2. On the other hand, if a certain number of normal waveforms are acquired and averaged, the characteristics of the waveform are smoothed by the averaging, and the characteristics of the waveform that should be judged pass / fail are lost, and the average waveform becomes a truly average waveform. May not be.

  Further, in the method of Patent Document 1, when a normal waveform includes a portion whose value changes suddenly, that is, a rising portion or a falling portion, the standard deviation in such a portion is unstable. It tends to be a neighbor. Since the observation of the waveform is started by the start trigger, the time of the observed waveform can be treated as the same, but the physical phenomenon occurring in the actual inspection object is not the same time Not limited to this, there is a slight time lag. When the rising portion deviates back and forth in the time axis direction, the time lag appears in the standard deviation as a result of the observed value variation. However, this standard deviation is not actually a variation in observed values but a variation in time. Therefore, actually, the observed values do not vary so much, but an excessive threshold waveform is created when the time varies and varies. That is, the waveform variation in the time axis direction is statistically evaluated as the waveform variation in the observation value axis direction, and is reflected in the shift amount. It cannot be said that the upper and lower limits are shown correctly in the downstream part.

  Further, in order to say that the upper and lower limits of the threshold waveform created by the method proposed in Patent Document 2 are appropriate, it is not sufficient to obtain a certain number of normal waveforms, It is necessary to acquire a plurality of normal waveforms that vary appropriately with a certain degree of variation. But this is not easy. Further, in order to verify whether or not a plurality of normal waveforms that are appropriately varied with a certain degree of variation can be acquired, after all, a statistical evaluation such as the standard deviation in Patent Document 1 is required.

  The present invention has been made in view of the above, and obtains a truly average average waveform without losing the characteristics of the waveform due to the smoothing effect by averaging the observed waveform, and the time of the observed waveform. It is an object of the present invention to obtain a threshold waveform creating apparatus that creates a threshold waveform that takes into account true variations in which variations in the axial direction do not appear as variations in the observed value axial direction.

  In order to solve the above-described problems and achieve the object, the present invention is a threshold waveform creation device that creates a threshold waveform used to determine the normality of an observed waveform, and a feature point is obtained from the observed waveform. A feature extraction unit for extraction, a correlation evaluation unit for evaluating the feature points extracted by the feature extraction unit, and a statistic for performing statistical processing in the time axis direction and the observation value axis direction of the observed waveform based on the feature points And a threshold waveform creating unit that creates the threshold waveform based on statistical processing in the time axis direction and the observed value axis direction.

  According to the present invention, a truly average average waveform without losing the characteristics of the waveform is obtained by the smoothing effect by averaging of the observed waveform, and the variation in the time axis direction of the observed waveform is the observed value axis. A threshold waveform that takes into account the true variation that does not appear as a variation in direction can be obtained. By using such a threshold waveform, the waveform and the threshold waveform that are measured when judging the quality of the product without excess or deficiency There is an effect that can be compared.

FIG. 1 is a diagram showing a hardware configuration of a first embodiment of a threshold waveform generating apparatus according to the present invention. FIG. 2 is a diagram illustrating a configuration of a program that executes creation of a threshold waveform. FIG. 3 is a diagram showing a point where the waveform structure changes greatly. FIG. 4 is a diagram showing the simplest type of waveform angle change. FIG. 5 is a diagram showing a first type that requires evaluation of a group of angle changes. FIG. 6 is a diagram showing a second type that requires evaluation of a group of angle changes. FIG. 7 is a diagram schematically illustrating the relationship between corresponding feature points between different waveforms. FIG. 8 is a diagram showing the feature points extracted by the feature extraction unit and their correlation. FIG. 9 is a diagram showing an average waveform, a standard deviation of observed values, and a standard deviation of time in the first embodiment. FIG. 10 is a diagram showing a threshold waveform created by adding the standard deviation of the observed value and the standard deviation of the time to the average waveform calculated in the first embodiment. FIG. 11 is a diagram schematically illustrating an average waveform evaluation method in the threshold waveform creating unit. FIG. 12 is a diagram illustrating feature point extraction by adjusting the degree of smoothing according to the second embodiment. FIG. 13 is a diagram schematically illustrating the evaluation of the standard deviation in the observed value axis direction based on the observed value point group in the flat section. FIG. 14 is a diagram showing the configuration of the third embodiment of the threshold waveform generating apparatus according to the present invention. FIG. 15 is a diagram schematically showing a process of creating a threshold waveform for the abnormality occurrence section in the third embodiment. FIG. 16 is a diagram schematically illustrating a threshold waveform generation method disclosed in Patent Document 1. FIG. 17 is a diagram schematically illustrating the correspondence between a plurality of observed waveforms in the statistical processing of Patent Document 1. FIG. 18 is a diagram showing a plurality of observed waveforms, average waveforms, standard deviations, and threshold waveforms in the statistical processing of Patent Document 1.

  Embodiments of a threshold waveform generating apparatus and method according to the present invention will be described below in detail with reference to the drawings. Note that the present invention is not limited to the embodiments.

Embodiment 1 FIG.
First, in order to promote the understanding of the first embodiment, the waveform comparison method and the problems of Patent Document 1 will be described with reference to FIGS.

FIG. 16 is a diagram schematically illustrating a threshold waveform generation method disclosed in Patent Document 1.
<Creation of threshold waveform of Patent Document 1>
Assuming that a test signal is given to a product that is known to be a non-defective product and that the behavior of the product at that time is measured at a sampling interval Δt, the start of the test is detected as a trigger and the start time t1 is held. At the same time, sampling is started, the end of the inspection is also detected by trigger, and the end time tn is held and the sampling is stopped. As a result, the time required for one cycle is tn-t1, and the number of waveform points is n. Since the time required for the inspection is constant every time, waveform acquisition is started at the start trigger of one cycle, and after that, when the number of waveform points reaches n points, the same thing is true. . In this way, a normal waveform (not shown) is obtained. Similarly, a test signal is given to a product that is known to be defective, and waveform acquisition is performed to obtain an abnormal waveform (not shown).

  In this way, several normal and abnormal waveforms are acquired, and a threshold waveform is defined based on the acquired some normal and abnormal waveforms. For example, an average waveform m obtained by averaging several acquired normal waveforms is shifted upward to create an upper limit waveform h, and downward to create a lower limit waveform l. The value in the observation value axis direction at time t of the average waveform m is assumed to be displayed as m (t). Similarly, the value in the observation value axis direction of the upper limit waveform h is indicated as h (t), and the lower limit waveform l The value in the observed value axis direction is represented by l (t). It is confirmed that any of the acquired abnormal waveforms deviate from the created threshold waveform in the sampling intervals t1 to tn, and finally the upper limit waveform h and the lower limit waveform l are set as threshold waveforms having the upper limit and the lower limit. Stipulate. The upper limit waveform h and the lower limit waveform l of the defined threshold waveform also have n waveform points (sampling interval Δt).

<Determination of Measurement Waveform of Patent Document 1>
Similarly, a test signal is given to the product to be inspected, and the behavior at that time is measured at the sampling intervals t1 to tn and the sampling interval Δt. As a result, the value in the observed value axis direction at time t is It is assumed that a measurement waveform (not shown) displayed at f (t) is obtained. Whether or not the measurement waveform deviates from the threshold waveform is determined by comparing the value of each point of the measurement waveform with the value of each point of the threshold waveform. For example, when both the upper limit waveform h and the lower limit waveform l are set as threshold waveforms as shown in FIG. 16, the value of each point of the measurement waveform is smaller than the value of each point of the upper limit waveform h, and , Larger than the value of each point of the lower limit waveform l, that is, l (t1) <f (t1) <h (t1), l (t2) <f (t2) <h (t2),..., L (tn) < If it can be confirmed that f (tn) <h (tn), the measured waveform does not deviate from the threshold waveform, and it is determined that the product is a non-defective product.

  In the above example, the horizontal axis is time (the number of sampling points) and the vertical axis is the observed value. When the observed waveform is converted into a waveform of a different unit in this way, for example, if it is Fourier transform (FFT), The horizontal axis after conversion is the frequency, and the vertical axis is the magnitude (spectrum) of the frequency component. Even in this case, the determination is made in the same manner if a threshold waveform for the converted waveform is set in advance. .

  As described above, in the method of Patent Document 1 that inspects a product after manufacturing processing and determines the quality of the product, it is necessary to provide an inspection step after the manufacturing processing step, and accordingly, per product. The time required from the start of production to completion is also increased. In order to solve this, by measuring the control commands and manufacturing process status during the manufacturing process, and confirming that the resulting waveform does not deviate from the waveform during normal manufacturing process There is a method for judging the quality of a product. This method is sometimes called in-line inspection or on-machine inspection (on-machine verification).

  This waveform determination method may be the same as in the above-described example. That is, instead of giving a test signal to the manufactured product and measuring the behavior of the product at that time, the status of the manufacturing process (one cycle) of one product is measured, and the obtained waveform is preset. The upper limit waveform and the lower limit waveform may be compared.

  On the other hand, the amount by which the above-described average waveform m is shifted up and down needs to be determined empirically. If some normal waveforms and some abnormal waveforms can be acquired, these normal waveforms can be judged as good, and the shift amount can be adjusted so that these abnormal waveforms can be judged as negative. In order to say that the amount is appropriate, it is necessary to acquire and compare a certain number of normal waveforms and abnormal waveforms, and the work becomes complicated. In addition, since the production line is usually used to manufacture non-defective products, it is difficult to acquire a certain number of abnormal waveforms compared to obtaining a certain number of normal waveforms relatively easily. Such work is extremely difficult.

  Therefore, it is statistically excellent to adopt the method disclosed in Patent Document 1 and determine the shift amount of the average waveform m in consideration of the standard deviations of a plurality of normal waveforms. In the method disclosed in Patent Document 1, an excessive threshold waveform may be created. The problem of creating this excessive threshold waveform will be described with reference to FIGS.

  FIG. 17 is a diagram schematically illustrating the correspondence between a plurality of observed waveforms in the statistical processing of Patent Document 1. The waveforms w1 to w3 are all waveforms having the same shape, but are slightly shifted from each other in the time axis direction. The values in the observed value axis direction at time t of the waveforms w1 to w3 are displayed as f1 (t) to f3 (t), respectively. For example, the observed values at time tk are assumed to be f1 (tk) to f3 (tk), respectively. The average and standard deviation of the observation values at time tk where the waveforms w1 to w3 are flat are not so problematic, but the observation values f1 (tj) to f3 () at the time tj at which the protruding portions of the waveforms w1 to w3 are observed. In the average and standard deviation of tj), the deviation of the waveforms w1 to w3 in the time axis direction appears as the deviation of the observed value, which causes a problem that an excessive threshold waveform is created. For example, originally, the observed value f3 (tj) of the waveform w3 should be statistically evaluated in terms of the correspondence with the observed value f1 (ti), not the observed value f1 (tj) of the waveform w1. That's why.

として求められ、同様にしてその標準偏差ｖ（ｔｋ）も、

として求めることができる。 Based on such conventional statistical processing, the observed values of the waveforms w1 to wn at the time tk are respectively displayed as f1 (tk) to fn (tk), and the average waveform value is displayed as m (tk). In general,

Similarly, the standard deviation v (tk) is also calculated as

Can be obtained as

  FIG. 18 is a diagram showing a plurality of observed waveforms, average waveforms, standard deviations, and threshold waveforms in the statistical processing of Patent Document 1. Each of the observed waveforms w4 to w6 shown in FIG. 18A is a normal waveform having the same shape characteristics with respect to the unevenness, but varies slightly in both the observation value axis direction and the time axis direction. ing. FIG. 18B shows an average waveform m1 obtained by averaging the waveforms w4 to w6 at each time based on Equation 1, but in the average waveform m1, the waveform w4 to w6 has a waveform immediately after the rise. It can be seen that the convex portions that are overshoot portions are almost lost. Based on Equation 1, the average waveform m1 does not average the maximum values of the convex portions of the waveforms w4 to w6, but the average of the observation values of the waveforms w4 to w6 at time 4, and the waveform w4 to w4 at time 5. This is because the average processing is performed at each time like the average of the observation values of w6. Further, FIG. 18C shows a value obtained by multiplying the standard deviation calculated at each time based on Equation 2 by a specific value. This value is obtained at the rising and falling portions of the waveforms w4 to w6. Protrusively large value. If a threshold waveform is created by the method of Patent Document 1 using a value based on this standard deviation, as shown in FIG. 18 (d), an upper limit waveform h1 and a lower limit that are excessive threshold values at the rising and falling portions. It can be seen that a threshold waveform whose upper and lower limits are defined by the waveform l1 is created, and it cannot be said that the correct upper limit or lower limit is indicated.

  Next, based on the description with reference to FIGS. 16 to 18 described above, the threshold waveform generation device according to the present invention will be described. FIG. 1 is a diagram showing a hardware configuration of a first embodiment of a threshold waveform generating apparatus according to the present invention. The threshold waveform generation device 1 includes a microprocessor 2, a system bus 3, a storage memory 4, an input unit 5, a storage unit 6, and a display unit 7. The storage memory 4 and the storage unit 6 can store a program for creating a threshold waveform. The microprocessor 2 performs processing according to this program, and creates an average waveform and a threshold waveform from the normal waveform stored in the storage unit 6. The display unit 7 is used for confirming the created average waveform and threshold waveform. The threshold waveform creation device 1 may be a personal computer, and the normal waveform observed by the inspection device or the measurement device is held on this personal computer, and the average waveform and the threshold waveform created offline on this personal computer are transferred to the inspection device. use. In addition, the threshold waveform generation apparatus 1 may be an inspection apparatus. In this case, the inspection apparatus has a function of generating an average waveform and a threshold waveform, and the average waveform and the threshold waveform are generated based on a normal waveform observed in advance of the inspection. Based on this, a pass / fail judgment is made on the observed waveform to be inspected.

  FIG. 2 is a diagram illustrating a configuration of a program that executes creation of a threshold waveform. The threshold waveform generation program 11 is stored in the storage memory 4 or the storage unit 6, and includes a noise processing unit 12, a feature extraction unit 13, a correlation evaluation unit 14, a statistical processing unit 15, and a threshold waveform generation unit. 16.

  The threshold waveform creation program 11 first removes noise from a normal waveform observed in advance by the noise processing unit 12 as necessary. This is because a normal waveform observed for a non-defective product may contain noise, and this noise prevents a point that is not originally a feature point of the waveform from being extracted as a feature point. Next, the feature extraction unit 13 extracts feature points of the observed normal waveform. Then, the correlation evaluation unit 14 evaluates the feature points extracted by the feature extraction unit 13, and adopts the feature points in which correlation is recognized among a plurality of normal waveforms as the true feature points. The statistical processing unit 15 calculates the average of the observed values of the true feature points of each normal waveform, and calculates the average of time. Here, the average calculation of observation values is a kind of statistical processing in the observation value axis direction, and the average calculation of time is a kind of statistical processing in the time axis direction. In addition, the statistical processing unit 15 calculates the average of the observed value and the time by assuming that the interval between the feature points corresponds uniformly, and generates a threshold waveform from the average waveform created from these calculation results One normal waveform is used as the basis for this. Further, the statistical processing unit 15 calculates the standard deviation of the observed value and the time for the corresponding point when the average of the observed value and the time is calculated. The standard deviation calculation of the observation value is a kind of statistical processing in the observation value axis direction, and the time standard deviation calculation is a kind of statistical processing in the time axis direction. Finally, the threshold waveform creating unit 16 creates a threshold waveform (upper limit waveform and lower limit waveform) in which the average value obtained by the statistical processing unit 15 is added with the observed value and the standard deviation of time. If necessary, compare multiple normal waveforms observed in advance with the extracted feature points, average waveform, or threshold waveform to determine whether the creation of the average waveform or threshold waveform is appropriate, The parameters related to the creation of the average waveform and threshold waveform are adjusted, and the average waveform and threshold waveform are created again.

  The noise processing unit 12 performs noise removal using a low-pass filter such as a moving average filter or a scale space filter. At this time, the degree of noise that can be removed can be adjusted by a time constant in the case of a moving average filter and by a scale (spread of a Gaussian function that is a basis function of convolution integral) in the case of a scale space filter. That is, the noise processing unit 12 can perform noise removal by adjusting the removal parameter 21 such as a time constant and a scale.

  While noise is removed by this, the feature point of the observed waveform may be lost due to the smoothing effect, so noise removal with an excessive removal parameter 21 is not preferable. For example, in the case of a moving average filter, a very large time constant should not be set and should be kept to a minimum. However, in this case, noise cannot be completely removed, and it is impossible to completely prevent a point that is not a characteristic point of a waveform from being extracted as a characteristic point. In order to cope with this, the correlation evaluation unit 14 is provided in the subsequent stage. Even so, the minimum noise removal is performed so that the feature point is excessively extracted and the original feature point is evaluated as having a correlation and is not adopted as the true feature point. Should be done.

  The feature extraction unit 13 extracts a plurality of feature points on the waveform structure from each of a plurality of observed normal waveforms. The feature point on the waveform structure is, for example, a point where the waveform structure changes greatly, and a value indicating the degree of change of the waveform exceeds a predetermined threshold. This point will be described with reference to FIGS.

  FIG. 3 is a diagram showing a point where the waveform structure changes greatly. FIG. 3 shows one observed waveform, and when the curvature of the waveform is adopted as a value indicating the degree of change of the waveform, the curvature exceeds a predetermined threshold at sampling points P1 to P6. Sampling points P1 to P6 are points where the waveform structure changes greatly, that is, characteristic points on the waveform structure. In addition, since the waveform is formed by connecting sampling points with line segments, the angle formed by the line segment is evaluated as a substitute for the curvature, and a point where the angle exceeds a predetermined threshold is also used as a characteristic point on the structure of the waveform. Good.

  When the feature extraction unit 13 extracts feature points, it is necessary to set the aspect ratio parameter 22 in addition to setting the feature reference parameter 23 serving as the predetermined threshold. Since the horizontal axis of the waveform is time, the vertical axis is the observed value and the unit is different, so the value of curvature and angle of the waveform also changes depending on how much the observed value is with respect to time. It is.

  Note that the determination of the point where the waveform structure changes significantly requires the waveform to be evaluated globally. This is because if it is determined with a threshold value only from the vicinity of the target sampling point, a portion that should be extracted as a feature point of the waveform may not be extracted as a feature point.

FIG. 4 is a diagram showing the simplest type of waveform angle change. FIG. 4 shows a waveform composed of sampling points P7 to P14. When an angle change is adopted as a value indicating the degree of change of the waveform and a feature point is extracted by determining the threshold value of the angle change, the angle change of the waveform is defined as, for example, angle changes Δθ ₁ and Δθ _2. it can. The angle change Δθ ₁ is an angle change of the sampling point P10 and corresponds to an inclination angle of the line segment connecting the sampling points P10 and P11 with respect to the line segment connecting the sampling points P9 and P10. Further, the angle change Δθ ₂ is an angle change of the sampling point P11 and corresponds to an inclination angle of the line segment connecting the sampling points P11 and P12 with respect to the line segment connecting the sampling points P10 and P11. As a feature reference parameter 23, the reference angle theta ₁ and +50 degrees as a criterion for determining a feature point has a concave shape, and the reference angle theta ₂ as a reference for determining a feature point that is a convex -50 ° If set, the sampling point P10 where the angle change Δθ ₁ makes +60 degrees is extracted as a feature point where the waveform is locally concave, and the point where the angle change Δθ ₂ makes -60 degrees is local in the waveform. Are extracted as feature points having a convex shape. However, even if the features of the concavo-convex shape have the same waveform, the concavo-convex shape may not be extracted as a feature point depending on how the sampling is performed.

FIG. 5 is a diagram showing a first type that requires evaluation of a group of angle changes. The waveform shown in FIG. 5 is composed of sampling points P7 to P14 as in the waveform shown in FIG. 4 and has the same feature of the uneven shape, but the angle change Δθ at any of the sampling points P9 and P10. ₃ and Δθ ₁ do not exceed the reference angle θ ₁ shown in FIG. 4, and none of the sampling points P9 and P10 is extracted as a feature point having a concave shape. The angle change Δθ ₃ is an angle change of the sampling point P9, and corresponds to an inclination angle of the line segment connecting the sampling points P9 and P10 with respect to the line segment connecting the sampling points P8 and P9.

FIG. 6 is a diagram showing a second type that requires evaluation of a group of angle changes. The waveform shown in FIG. 6 is composed of sampling points P7 to P14 as in the waveform shown in FIG. 4 and has the same feature of the uneven shape, but the angle change Δθ at any of the sampling points P11 and P12. ₂ and Δθ ₄ do not exceed the reference angle θ ₂ shown in FIG. 4, and none of the sampling points P11 and P12 is extracted as a feature point having a convex shape. Note that the angle change Δθ ₄ is an angle change of the sampling point P12 and corresponds to an inclination angle of the line segment connecting the sampling points P12 and P13 with respect to the line segment connecting the sampling points P11 and P12.

In FIGS. 5 and 6, in order to prevent such a situation, a group of points that change continuously in the same direction is handled as one group, and the sum of these changes in angle is regarded as a group of changes in angle, which is set as a threshold value. Should be judged. For example, in the case of the waveform of FIG. 5, the sampling points P9 and P10 are handled as a group, and it is determined that the angle change Δθ ₁ ′ (= angle change Δθ ₃ + angle change Δθ ₁ ) of the group exceeds the reference angle θ _1. In such a case, the center point of the point group (sampling points P9, P10) may be treated as a feature point having a concave shape. Similarly, in the case of FIG. 6, the sampling points P11 and P12 are handled as a group, and it is determined that the angle change Δθ ₂ ′ (= angle change Δθ ₂ + angle change Δθ ₄ ) of the group exceeds the reference angle θ _2. In this case, the middle point of the point group (sampling points P11 and P12) may be handled as a feature point having a convex shape.

  FIG. 7 is a diagram schematically illustrating the relationship between corresponding feature points between different waveforms. FIG. 7 shows two waveforms W1 and W2, and the observed values of the waveforms W1 and W2 at time T are indicated by F1 (T) and F2 (T), respectively. A sampling point P15 corresponding to the observation value F1 (T2) on the waveform W1 and a sampling point P21 corresponding to the observation value F2 (T3) on the waveform W2 are the start points of the rise, and feature points corresponding to each other at different times Extracted as A sampling point P16 corresponding to the observation value F1 (T3) on the waveform W1 and a sampling point P22 corresponding to the observation value F2 (T4) on the waveform W2 are the end points of the rise, and correspond to each other at different times. Extracted as feature points. Next, the sampling point P19 corresponding to the observation value F1 (T7) on the waveform W1 and the sampling point P24 corresponding to the observation value F2 (T6) on the waveform W2 are the start points of the fall, and are mutually different at different times. It is extracted as a corresponding feature point. The sampling point P20 corresponding to the observation value F1 (T8) on the waveform W1 and the sampling point P25 corresponding to the observation value F2 (T7) on the waveform W2 are the end points of the fall and correspond to each other at different times. Are extracted as feature points. Further, assuming that the section between each feature point corresponds uniformly, the sampling point P18 corresponding to the observed value F1 (T5) on the waveform W1 is sampled corresponding to the observed value F2 (T5) on the waveform W2. Assume that it corresponds to the point P23. Note that the sampling point P17 corresponding to the observed value F1 (T4) on the waveform W1 presumes that there is a corresponding point at the midpoint between the sampling points P22 and P23 on the waveform W2.

で求められ、その標準偏差は、観測値軸方向の標準偏差Ｖｆ（Ｔｉ）および時間軸方向の標準偏差Ｖｔ（Ｔｉ）として、

でそれぞれ求められる。 Therefore, by extracting the feature points of the normal waveform by the feature extraction unit 13, it is possible to determine corresponding points among a plurality of normal waveforms as described in FIG. Therefore, the threshold waveform creation program enables statistical evaluation such as average and standard deviation in both the observation value axis direction and the time axis direction. For example, in the case of FIG. 7, the average value M (Ti) of the observed values at the sampling points P15 and P21, which are the start points of the rising edges of the waveforms W1 and W2, is

The standard deviation is obtained as standard deviation Vf (Ti) in the observed value axis direction and standard deviation Vt (Ti) in the time axis direction.

Each is required.

  The correlation evaluation unit 14 adopts, as true feature points, features that are correlated with respect to feature points extracted from a plurality of normal waveforms, and the true features so that the same group includes a plurality of feature points having the same correlation. Group points into multiple groups. This is to prevent a point that is not a characteristic point of the waveform from being extracted as a characteristic point due to the influence of noise. Since the plurality of normal waveforms have the same waveform characteristics, feature points are extracted at approximately the same position from each normal waveform. It can be understood that the feature points extracted from only some of the waveforms exceed the reference value for feature determination due to the influence of noise. This point will be described with reference to FIG.

  FIG. 8 is a diagram showing the feature points extracted by the feature extraction unit and their correlation. 8A shows the waveform W3 and the feature points P26 to P30 of the waveform W3, FIG. 8B shows the waveform W4 and the feature points P31 to P36 of the waveform W4, and FIG. 8C shows the waveform W5 and the waveform. Characteristic points P37 to P42 of W5 are shown, and FIG. 8D shows a state in which only these characteristic points P26 to P42 are plotted. It can be seen that the concave feature points P26, P31, and P37 extracted from the waveforms W3 to W5 exist close to each other in the vicinity of time 3 to 5 in FIG. In this way, it is understood that the same kind of feature points (same concave feature points or same convex feature points) that are extracted from any of the waveforms W3 to W5 and are close to each other within a certain range have a correlation. Therefore, the correlation evaluation unit 14 adopts the true feature point. Further, the correlation evaluation unit 14 groups the feature points P26, P31, and P37 in which the same correlation is recognized among the plurality of waveforms W3 to W5, and adopts them as point groups corresponding to each other. A threshold value that is a criterion for determining how close a range is to be close is determined or adjusted as a correlation parameter 24. For example, examples of the correlation parameter 24 include a proximity range in the time axis direction and a proximity range in the observation value axis direction. Thereby, the correlation evaluation unit 14 determines that the convex feature points P27, P32, and P38 extracted from the waveforms W3 to W5 in the vicinity of the time 4 to 6 are the same kind of feature points that are close to each other and have a correlation. Therefore, the concave feature points P28, P33, and P39, which are adopted as true feature points and grouped as a point group where the same correlation is recognized and extracted from the waveforms W3 to W5 in the vicinity of time 5 to 8, are Since it is determined that there is a correlation between the same kind of feature points that are close to each other, they are adopted as true feature points and are grouped as a point group in which the same correlation is recognized. The convex feature points P29, P35, and P41 extracted from W5 are the same type of feature points that are close to each other and are determined to be correlated, so that they are true feature points. The concave feature points P30, P36, and P42 extracted from the waveforms W3 to W5 in the vicinity of the time 25 to 27 are grouped as point groups that are adopted and have the same correlation, and are similar feature points that are close to each other. Since it is determined that there is a correlation, it is adopted as a true feature point and is grouped as a point group in which the same correlation is recognized. In addition, concave feature points P26, P31, and P37 around the time 3 to 5 and concave feature points P28, P33, and P39 around the time 5 to 8 are the same kind of concave feature points, and the times are close to each other. However, the observed values are not close to each other, and the correlation evaluation unit 14 does not determine that there is a correlation. That is, the feature points P26, P31, P37, P28, P33, and P39 are not grouped as a point group in which the same correlation is recognized. On the other hand, the convex feature point P34 extracted from only the waveform W4 in the vicinity of time 15 exists close to the convex feature points P27, P29, P38, and P41 extracted from the other waveforms W3 and W5. Since there is nothing, it cannot be adopted as a true feature point. The concave feature point P40 extracted from only the waveform W5 in the vicinity of the time 16 is also close to the concave feature points P26, P28, P30, P31, P33, and P36 extracted from the other waveforms W3 and W4. Since there is nothing, it cannot be adopted as a true feature point.

  The statistical processing unit 15 creates an average waveform by calculating the average of the observed values and the time of the feature points of each normal waveform that are adopted as true feature points by the correlation evaluation unit 14 and grouped for each group. To do. In addition, the interval between feature points is assumed to correspond uniformly, the corresponding points are determined, and the average of the observed value and the time is calculated from the corresponding points of each normal waveform. The true feature point has a point corresponding to the sampling point in each normal waveform, but the point corresponding to the sampling point on one normal waveform is the other normal waveform in the interval between the feature points. Since there may be a point on the line segment that does not exist as an upper sampling point, a point on such a line segment may be calculated, or a sampling point existing in the vicinity of the corresponding point may be calculated. Alternatively, it may be approximated instead of

  If the corresponding points between the normal waveforms are determined in this way, an average waveform is obtained based on Equations 3 and 4, and becomes a normal waveform that is the basis for creating the threshold waveform. The statistical processing unit 15 further calculates standard deviations in the observation value axis direction and the time axis direction based on the equations 5 and 6, respectively. Each normal waveform is composed of a group of n sampling points, and the time of each sampling point is an integer value when the time is a sampling number, but the time of each point of the obtained average waveform is the average calculation time. As a result, it is usually not an integer value. As shown in FIG. 16, in order to compare waveforms, the average waveform and the threshold waveform are also composed of n sampling point groups, and the time needs to be matched. Therefore, the result of the average operation may be rounded by rounding or the like to obtain an integer, which may be used as an average waveform, or a line segment consisting of a point group indicated by decimal coordinate values is interpolated and indicated by integer coordinate values. It is also possible to calculate a point group to be averaged.

  FIG. 9 is a diagram showing an average waveform, a standard deviation of observed values, and a standard deviation of time in the first embodiment. FIG. 9A shows, for each point group in which the same correlation is recognized from the true feature points of each normal waveform adopted as the true feature points shown in FIG. The average waveform M1 created by calculating the average of the observed values and the time based on Equations 3 and 4 is shown. The feature point P43 of the average waveform M1 indicates the average value of the observed values and the average value of the time of the feature points P26, P31, and P37 in FIG. 8D, and from the feature points P26, P31, and P37 where the same correlation is recognized. It is a point corresponding to a point group. Similarly, the feature point P44 of the average waveform M1 indicates the average value and the average value of the observed values of the feature points P27, P32, and P38 in FIG. 8D, and feature points P27 and P32 where the same correlation is recognized. , P38, the point corresponding to the point group. The feature point P45 indicates an average value of observed values and an average value of time of the feature points P28, P33, and P39 in FIG. 8D, and is a point group including feature points P28, P33, and P39 that have the same correlation. It is a corresponding point. The feature point P46 indicates the average value and the average value of the observed values of the feature points P29, P35, and P41 in FIG. 8D, and is a point group including the feature points P29, P35, and P41 where the same correlation is recognized. It is a corresponding point. The feature point P47 indicates an average value of observation values and an average value of time of the feature points P30, P36, and P42 in FIG. 8D, and is a point group including feature points P30, P36, and P42 in which the same correlation is recognized. It is a corresponding point. FIG. 9B shows, for each point group in which the same correlation is recognized from the true feature points of the respective normal waveforms adopted as the true feature points shown in FIG. The standard deviation of the observed value calculated based on Equation 5 is shown. FIG. 9C shows, for each point group in which the same correlation is recognized from the true feature points of the respective normal waveforms adopted as the true feature points shown in FIG. The standard deviation of the time calculated based on Equation 6 is shown. The statistical processing values shown in FIGS. 9A to 9C are the ones in which the time lag of the waveform in the conventional method shown in FIG. It is evaluated as the variation in the observed value axis direction and the time axis direction.

  The threshold waveform creating unit 16 is the same as that shown in FIG. 8D corresponding to each of the feature points P43 to P47 with respect to each of the feature points P43 to P47 of the average waveform M1 shown in FIG. A threshold waveform (upper limit waveform and lower limit waveform) is created by taking into account the standard deviation of the observed values and the standard deviation of time calculated for each point group where correlation is recognized, according to a predetermined tolerance. The predetermined tolerance refers to adding a certain amount of the standard deviation of the observation location or time. The certain tolerance is set or adjusted by the tolerance parameter 25. This point will be described with reference to FIGS. 10 and 11.

  FIG. 10 is a diagram showing a threshold waveform created by adding the standard deviation of the observed value and the standard deviation of the time to the average waveform calculated in the first embodiment. FIG. 11 is a diagram schematically illustrating an average waveform evaluation method in the threshold waveform creating unit. FIG. 10A shows an ellipse composed of the standard deviation of the observed values and the standard deviation of the time at the characteristic points P43 to P47 of the average waveform M1. Here, an ellipse is shown when the observed value and three times the standard deviation of time are set as the tolerance parameter 25. When the threshold waveform creating unit 16 evaluates the average waveform M1 from the viewpoint of the ellipse shown in FIG. 10 (a), as shown in FIG. 10 (b), the waveform constituting the outer edge of the ellipse group is set to the upper limit waveform H1 and It can be defined as the lower limit waveform L1.

  Here, as an example of a method of constructing the upper limit waveform H1 and the lower limit waveform L1 constituting the outer edge of the ellipse group, a space surrounded by the ellipse group (in FIG. 11, the ellipse C1 shown at each time (FIG. 10 (a ), The value of three times the standard deviation of the observed value corresponding to the time T10 on the average waveform corresponding to the average waveform M1 is taken as the length of the short axis A1, and the standard deviation of the time corresponding to the time T10 on the average waveform is 3 An ellipse having a double value as the length of the major axis A2), an ellipse C2 (a value that is three times the standard deviation of the observed value corresponding to the time T11 on the average waveform) is the length of the minor axis A3, and a time T11 on the average waveform The ellipse C3 (the ellipse having a value of three times the standard deviation of the time corresponding to the length of the major axis A4) and the ellipse C3 (three times the standard deviation of the observed value corresponding to the time T12 on the average waveform) A standard deviation of time corresponding to time T12 on the average waveform. A range surrounded by an ellipse having a value of 3 times the short axis A6) is evaluated as a space to be judged as good, and an average waveform is evaluated, and an upper limit waveform and a lower limit waveform as threshold waveforms are determined from the average waveform. (As shown in FIG. 11, the lengths of the corresponding short axes A1, A3, A6 or long axes A2, A4, A5 are adjusted for the points on the average waveform corresponding to the respective times T10 to T12. The average waveform is shifted in the observed value axis direction and the time axis direction). For example, in the ellipse C1 in FIG. 11, the length of the short axis A1 is added to the observation value corresponding to the time T10 of the average waveform, and the length of the long axis A2 is subtracted from the time corresponding to the time T10 of the average waveform. Then, it becomes a point on the upper limit waveform. Similarly, in the ellipse C1, when the length of the minor axis A1 is subtracted from the observed value corresponding to the time T10 of the average waveform, and the length of the major axis A2 is added to the time corresponding to the time T10 of the average waveform, the lower limit waveform This is the top point. That is, in the method disclosed in Patent Document 2, a threshold waveform is created with a space surrounded by a plurality of observed normal waveforms as a space to be determined as good. The first embodiment In the method, a threshold waveform is created with a space surrounded by an ellipse group composed of the standard deviation of the observed values and the standard deviation of time for each point of the average waveform as a space to be determined as good.

  In the process up to this point, parameters to be adjusted include a removal parameter 21, an aspect ratio parameter 22, a feature criterion parameter 23, a correlation parameter 24, and a tolerance parameter 25. The average waveform M1 and the threshold waveform (the upper limit waveform H1 and the lower limit waveform L1) are also different. Therefore, the extracted feature points P43 to P47, average waveform M1, threshold waveform (upper limit waveform H1 and lower limit waveform L1) are compared with a plurality of observed normal waveforms W3 to W5 to determine whether or not they are appropriate. This parameter can be adjusted as necessary, and the average waveform M1 and the threshold waveform can be created again. For example, as shown in FIG. 10C, the parameters are adjusted so that the observed normal waveforms W3 to W5 are within the obtained threshold waveforms (upper limit waveform H1 and lower limit waveform L1). In setting or adjusting these parameters, a single constant value may be set and applied throughout the entire waveform, or a different constant value may be set and finely adjusted for each time.

  As described above, according to the first embodiment, it is possible to evaluate the variation of the obtained plurality of normal waveforms W3 to W5 in both the observation value axis direction and the time axis direction. The average normal waveform M1 can be obtained without losing the characteristics of the waveform due to the effect, and the standard deviation in the observed value axis direction can be obtained without the time variation appearing as the observed value variation. The amount of shift can be determined by the standard deviation in the time and time direction, and threshold waveforms (upper limit waveform H1 and lower limit waveform L1) taking into account true variation can be generated. Then, by using such a threshold waveform, waveform comparison without excess or deficiency can be realized.

  The pass / fail judgment based on the threshold waveform according to the first embodiment can be used not only for product inspection, but also for abnormality detection of the monitoring control system such as an alarm for abnormalities in social infrastructure and an alarm for daily power consumption. In the conventional monitoring and control system, only standard determination based on the upper limit value, the lower limit value, the upper upper limit value, and the lower lower limit value was prepared as a criterion for detecting an abnormality. For example, a threshold waveform can be created from a normal power pattern among the daily power consumption patterns, and more precise energy saving judgment can be made, and wastefulness can be prevented by warning. Even in this case, since the start time and the end time of power consumption may be shifted from day to day, the threshold waveform created by the first embodiment is more appropriate than the conventional case, compared to the judgment by the threshold waveform created by the conventional method. An alarm can be obtained by judgment.

Embodiment 2. FIG.
In the first embodiment, it is necessary to observe a certain number of normal waveforms in order for the evaluation values obtained by statistical processing such as average and standard deviation to be sufficiently valid. Therefore, the second embodiment makes it possible to create an average waveform or a threshold waveform by estimating an evaluation value by statistical processing such as an average or standard deviation without observing such a number of normal waveforms. The method will be described. In the second embodiment, the threshold waveform creation program 11 creates an average waveform or a threshold waveform by estimating an evaluation value by statistical processing such as an average or standard deviation, taking the case where one normal waveform is observed as an example. Will be described.

  In the first embodiment, noise is removed in the noise processing unit 12 in order to prevent a point that is not a characteristic point of a waveform from being extracted as a characteristic point due to noise superimposed on the waveform, but by the smoothing effect. In order to prevent the feature points of the waveform from being lost, noise removal by the minimum removal parameter 21 is limited, and the correlation evaluation unit 14 associates only correlated feature points as true feature points. However, if there is only one normal waveform, the correlation evaluation unit 14 cannot evaluate the correlation as in the first embodiment. Therefore, in the second embodiment, the noise is removed by the large removal parameter 21 to some extent, and the feature reference parameter 23 at the time of extracting the feature point is set to be gentle. If the degree of smoothing is gradually increased, noise that is a high-frequency component is removed first, but the characteristic structure of the original waveform is gradually lost thereafter, so that the noise can be removed to a large extent. While using the removal parameter 21, the feature of the waveform that has not been lost is tried to be extracted by lowering the threshold value of the feature reference parameter 23.

  FIG. 12 is a diagram illustrating feature point extraction by adjusting the degree of smoothing according to the second embodiment. One normal waveform W6 observed includes noise. In the second embodiment, scale space filtering is used as a noise removal method. If the scale s in the scale space filtering is increased, only the rough structure of the waveform W6 is left. Therefore, the scale s is adjusted as a denoising parameter. Noise is gradually removed from each waveform W61, waveform W62, and waveform W63 after scale interval filtering when scale s = 1, scale s = 2, and scale s = 3 for one waveform W6. When the feature extraction unit 13 extracts feature points for the waveform W63 corresponding to the scale s = 3, convex feature points P49 and P51 and concave feature points P48, P50, and P52 are extracted, and the correlation evaluation unit 14, feature points P48 to P52 are adopted as true feature points. Scale space filtering is disclosed in non-patent literature (scale space filtering of periodic waveforms (The IEICE Transactions D Vol.J73-D2 No.4 pp.544-552, dated April 25, 1990)). As shown, it is guaranteed that the structure of the waveform W6 is monotonously lost as the scale s is increased.

  If there is only one normal waveform, the correlation cannot be evaluated, and neither the average nor the standard deviation can be evaluated. Therefore, in the correlation evaluation unit 14, the feature points P53 to P57 on the waveform W6 before the noise removal of the feature points P48 to P52 extracted with respect to the waveform W63 from which noise is removed with the scale s = 3 are used as the waveform W6. Is adopted as a true feature point, and a flat section is specified among the sections between the true feature points P53 to P57. That is, the correlation evaluation unit 14 specifies a flat section of the waveform W6 based on the feature points P48 to P52. The correlation parameter 24 in the second embodiment sets or adjusts a threshold value as a reference for determining how much the slope is flat. For example, if two observed values of the feature points P53 to P57 are within the threshold range, the section between the two feature points is determined to be flat, or two positions of the feature points P53 to P57 are determined. For example, a straight line approximation is performed on a point group between points by a least square method or the like, and if the slope of the straight line is within a threshold range, it is determined that the point is flat. Then, the statistical processing unit 15 handles the one normal waveform W6 as an average waveform.

で求められるから、標準偏差Ｖｆ_ｊｋは、

で求められる。この標準偏差Ｖｆ_ｊｋを、平坦区間Ｔ_ｊｋの各点における観測値軸方向の標準偏差として取り扱う。波形Ｗ６が平坦となっている部分は、理想的には平坦な直線となるべきところ、重畳するノイズのほか、観測対象の物理現象自体のゆらぎによって、観測値にばらつきが出ているのであるから、複数波形の観測値のばらつきを統計評価する代わりに、同一波形における平坦部分の観測値のばらつきを統計評価しても、その趣旨に逸脱しない。そして、算出された標準偏差Ｖｆ_ｊｋを、平坦区間Ｔ_ｊｋの各サンプリング点における時間軸方向の標準偏差として擬制し、また、特徴点Ｐ５３〜Ｐ５７の間のうち平坦区間Ｔ_ｊｋを除く他の区間の各サンプリング点における観測値軸方向の標準偏差や時間軸方向の標準偏差として擬制して取り扱う。正常波形が波形Ｗ６の１つである以上、平坦区間Ｔ_ｊｋの観測値軸方向以外のばらつきについては、その趣旨に逸脱しない妥当な算出方法が見いだせないため、ばらつきは波形Ｗ６の全体をとおして一様であるものと擬制するのである。 FIG. 13 is a diagram schematically illustrating the evaluation of the standard deviation in the observed value axis direction based on the observed value point group in the flat section. The statistical processing unit 15 targets the point group belonging to the flat section T _jk specified by the correlation evaluation unit 14 as a target, and as a statistical process only in the observation value axis direction, a standard that is a variation in observation values corresponding to the point group The deviation Vf is calculated. That is, when the observed value at time T of the average waveform (normal waveform W6) is displayed as F (T) and the observed value in the flat section T _jk is from F (Tj) to F (Tk), the average value M _jk is

_Therefore , the standard deviation Vf _jk is

Is required. This standard deviation Vf _jk is treated as a standard deviation in the observation value axis direction at each point of the flat section T _jk . The portion where the waveform W6 is flat should ideally be a flat straight line, because the observed value varies due to fluctuations of the physical phenomenon itself to be observed in addition to the superimposed noise. Instead of statistically evaluating the dispersion of observed values of a plurality of waveforms, statistically evaluating the dispersion of observed values of flat portions in the same waveform does not depart from the spirit. Then, the standard deviation _{Vf jk} calculated, and constructive as the standard deviation of the time axis at each sampling point of the flat section _{T jk,} also other sections except the flat sections _{T jk} of between feature points P53~P57 Are treated as a standard deviation in the observation value axis direction and a standard deviation in the time axis direction at each sampling point. As long as the normal waveform is one of the waveforms W6, for variations other than the observed value axis direction of the flat section _Tjk, an appropriate calculation method that does not deviate from the purpose cannot be found. It pretends to be uniform.

  Thus, since the standard deviation of the observed value and the standard deviation of time for the waveform W6 as an average waveform and each point thereof can be obtained, the threshold waveform creating unit 16 performs the same as in the first embodiment. Thus, a threshold waveform can be created. In addition, when the obtained normal waveform is 2 or more but is a small number and the evaluation value by the statistical processing such as the average or the standard deviation is not sufficiently valid, it is a small number but a plurality of The statistic (average and standard deviation) calculated from the normal waveform according to the first embodiment and the statistic (average and standard deviation) estimated according to the second embodiment for each normal waveform are averaged or weighted average. By doing so, the threshold waveform can be created after making one statistic.

  As described above, according to the second embodiment, even if one or a small number of normal waveforms are obtained, the statistics evaluated in the first embodiment are replaced by the estimation based on the estimation. The threshold waveform can be created by the same method as in FIG.

Embodiment 3 FIG.
In the first and second embodiments, the threshold waveform is created from the normal waveform. However, when the cause of the abnormality is known in advance and the abnormal waveform due to the same kind of cause can be observed, It is not necessary to set a threshold waveform for the entire observed waveform, and it may be sufficient to set a threshold waveform in the vicinity where the abnormality occurs. In the third embodiment, the setting of such a threshold waveform will be described.

  FIG. 14 is a diagram showing the configuration of the third embodiment of the threshold waveform generating apparatus according to the present invention. In the third embodiment, the threshold waveform creating program 11 is newly provided with an abnormality occurrence section extracting unit 33, whereby the threshold waveform creating unit 16 sets a threshold waveform only in the vicinity where an abnormality occurs in the observed waveform. It has a function to do. FIG. 15 is a diagram schematically showing a threshold waveform creation process for an abnormality occurrence section in the third embodiment. FIG. 15A shows an average calculated based on the threshold waveform creation program 11 according to the third embodiment. The waveform M2 is shown, FIG.15 (b) shows abnormality generation area S2, and FIG.15 (c) shows the threshold waveform (upper limit waveform H2 and lower limit waveform L2) produced only with respect to abnormality generation area S2.

  In the third embodiment, the threshold waveform creation program 11 performs noise removal on the normal waveform 31 or abnormal waveform 32 observed in advance of the inspection as necessary by the noise processing unit 12, and the feature extraction unit The feature points are extracted at 13, and the correlation evaluation unit 14 and the statistical processing unit 15 perform correlation evaluation and statistical processing to calculate the average waveform M2 and the standard deviation. That is, when a plurality of normal waveforms 31 can be observed, the correlation is evaluated and true feature points are extracted in the same manner as in the first embodiment, and when only one normal waveform can be used, the second embodiment Similarly, feature points are extracted. When a plurality of abnormal waveforms 32 can be observed, feature points that are mutually correlated between the plurality of abnormal waveforms 32 are extracted as true feature points in the same manner as in the first embodiment. When only the waveform 32 can be used, feature points are extracted as in the second embodiment.

  Next, the abnormality occurrence section extraction unit 33 evaluates the correlation between the feature points of the normal waveform 31 and the feature points of the abnormal waveform 32, and is constituted by feature points having no correlation as shown in FIG. The section to be processed is recognized as the abnormality occurrence section S2. At this time, the extent to which the abnormality occurrence section S2 is recognized is adjusted by the abnormality reference parameter 34. For example, as shown in FIG. 15B, the positive and negative directions of the time axis with respect to the section S1 defined by the characteristic points P58 to P61 of the abnormal waveform 32 that have no correlation with the characteristic points of the normal waveform 31 Further, it may be considered that the abnormality occurrence section S2 is recognized with a width obtained by adding a value of three times the standard deviation V of the time corresponding to the feature points P58 to P61 of the abnormal waveform 32. When only one abnormal waveform 32 can be used, it may be based on the standard deviation simulated by the second embodiment. However, the standard deviation in the time axis direction at the feature point of the section of the normal waveform 31 corresponding to the section S1 is used. May be based. In other words, based on the feature points of the normal waveform 31 and the feature points of the abnormal waveform 32 that are correlated with each other, if it is assumed that each point corresponds uniformly, the abnormal waveform 32 that has no correlation with the feature points of the normal waveform 31. The characteristic points P58 to P61 are based on the standard deviation of the time of the corresponding points on the normal waveform 31.

  Finally, as shown in FIG. 15C, only for the abnormality occurrence section S2 obtained in this way, the threshold waveform creating unit 16 uses the average waveform M2 and the standard deviation calculated by the statistical processing unit 15 to set the threshold value. Waveforms (upper limit waveform H2 and lower limit waveform L2) are created.

  According to the third embodiment, it is not necessary to determine the upper and lower limits for the entire observed waveform, and the processing time for the determination is shortened. As a result of the need for high production efficiency, the time from one product processing to the next one product processing is shortened, and it is beneficial when judgment processing must be completed within that short time. Bring.

  As described above, the threshold waveform creating apparatus according to the present invention can obtain a threshold waveform that takes into account the true variation in which the variation in the time axis direction of the observed waveform does not appear as the variation in the observed value axis direction. This is useful in that such a threshold waveform is used.

DESCRIPTION OF SYMBOLS 1 Threshold waveform creation apparatus, 2 Microprocessor, 3 System bus, 4 Storage memory, 5 Input part, 6 Storage part, 7 Display part, 11 Threshold waveform creation program, 12 Noise processing part, 13 Feature extraction part, 14 Correlation evaluation part , 15 Statistical processing unit, 16 Threshold waveform creation unit, 21 Removal parameter, 22 Aspect ratio parameter, 23 Feature standard parameter, 24 Correlation parameter, 25 Tolerance parameter, 31 Normal waveform, 32 Abnormal waveform, 33 Abnormality generation section extraction unit 34, abnormal reference parameters, P1 to P25 sampling points, P26 to P61 feature points, θ ₁ , θ ₂ reference angles, Δθ _{1 to} Δθ ₄ angular changes, Δθ ₁ ^′ , Δθ ₂ ^′ group of angular changes, W1 to W6 w1-w6 waveform, M1, M2, m, m1 average waveform, H1, H2, h, h1 upper limit waveform, L1, L2, l, l1 lower limit waveform, 1~C3 oval, A1, A3, A6 short axis, A2, A4, A5 long axis, s _{scale, T jk} flat sections, S1 segment, S2 abnormal interval.

In order to solve the above-described problems and achieve the object, the present invention is a threshold waveform creation device that creates a threshold waveform used to determine the normality of an observed waveform, and a feature point is obtained from the observed waveform. A feature extraction unit for extraction, a correlation evaluation unit for evaluating the feature points extracted by the feature extraction unit, and a statistic for performing statistical processing in the time axis direction and the observation value axis direction of the observed waveform based on the feature points A processing unit, and a threshold waveform generation unit that generates the threshold waveform based on statistical processing in the time axis direction and the observation value axis direction, and the feature extraction unit includes a plurality of observation waveforms. A plurality of feature points are extracted, and the correlation evaluation unit groups the feature points into a plurality of feature points so that the same group includes a plurality of feature points having the same correlation. The group And calculating the average of the observed values of the feature points and the average of the time to create an average waveform, and calculating the standard deviation of the observed values and the time, and the threshold waveform creating unit corresponds to the group for points on the average waveform, in consideration of the standard deviation, it characterized that you create the threshold waveform.

Claims (7)

A threshold waveform creating device that creates a threshold waveform used to determine the normality of an observed waveform,
A feature extraction unit for extracting feature points from the observed waveform;
A correlation evaluation unit for evaluating the feature points extracted by the feature extraction unit;
Based on the feature points, a statistical processing unit that performs statistical processing in the time axis direction and observation value axis direction of the observed waveform;
A threshold waveform creating unit that creates the threshold waveform based on statistical processing in the time axis direction and the observation value axis direction;
A threshold waveform generating device comprising: The feature extraction unit extracts a plurality of feature points from each of the plurality of observed waveforms,
The correlation evaluation unit groups the feature points into a plurality such that the same group includes a plurality of feature points having the same correlation,
The statistical processing unit calculates an average of the observed values of the feature points and an average of the time for each group as the statistical processing to create an average waveform, and calculates a standard deviation of the observed values and the time And
The threshold waveform creating device according to claim 1, wherein the threshold waveform creating unit creates the threshold waveform by adding the standard deviation to the points on the average waveform corresponding to the group. .   The threshold waveform creation unit adds or subtracts a value based on the observed value of the group and the standard deviation of the time corresponding to the point on the average waveform to obtain an upper limit waveform and a lower limit waveform as the threshold waveform. The threshold waveform creating device according to claim 2, wherein the threshold waveform creating device is created. The feature extraction unit extracts the feature points from one observed waveform,
The correlation evaluation unit identifies a flat section of the one observation waveform based on the feature point,
The statistical processing unit calculates a standard deviation of observed values corresponding to a plurality of point groups belonging to the flat section, and simulates the calculated standard deviation of the observed values as a standard deviation of time in the flat section. , As a standard deviation of observed values and time in other sections of the section excluding the flat section,
The threshold waveform creating apparatus according to claim 1, wherein the threshold waveform creating unit creates the threshold waveform by adding the observed value and the standard deviation of time to the one observed waveform. An abnormality occurrence section extraction unit that recognizes a section including a feature point having no correlation between a feature point of a normal waveform of the observed waveform and the feature point of an abnormal waveform of the observation waveform as an abnormality occurrence section,
The threshold waveform creating device according to claim 1, wherein the threshold waveform creating unit creates the threshold waveform only for the abnormality occurrence section.   6. The threshold waveform generation device according to claim 1, wherein the feature extraction unit extracts, as the feature point, a point where a value indicating a degree of change in the observed waveform exceeds a threshold value. .   The threshold waveform according to any one of claims 1 to 6, further comprising: a noise processing unit that removes noise from the observed waveform as preprocessing of the feature point extraction processing in the feature extraction unit. Creation device.

Images