JP5017678B2 - Signal inspection method and signal inspection module - Google Patents

Description

In the field of object state diagnosis and pattern recognition, the present invention comprehensively evaluates whether or not the signal of the object obtained by measurement has overflow, outlier and non-stationarity, and was measured. This is a method for inspecting the quality of a signal, and relates to a module for signal inspection that can be provided in order to prevent misdiagnosis and misjudgment in a diagnosis device, a pattern recognition device, and the like.

When measuring a signal for object state diagnosis or pattern recognition, the processing method differs depending on whether the measured signal is a steady signal or an unsteady signal. If an unsteady signal is input to a diagnostic device or pattern recognition device for processing a steady signal, an incorrect result is output.

In addition, the true signal emitted from an object such as diagnosis or pattern recognition is a steady signal, but the signal measured due to a measurement error or disturbance is an unsteady signal, and is further measured when an overflow occurs. The resulting signal is an incomplete signal that is not the true signal of the object. Even if such an unsteady signal or an inadequate signal is input to a diagnostic device or a pattern recognition device, an accurate result cannot be obtained.

Conventionally, various diagnostic devices and pattern recognition devices have been proposed in various fields, but the quality of the measured signal (measurement errors, processing defects, etc.) before performing the processing for diagnosis and pattern recognition Inspection method and module for the presence / absence) are not shown. In this regard, a particularly related past document is the prior application [Patent Document 1], and other related documents are, for example, [Patent Documents 2, 3, 4, 5, 6].
Japanese Patent Application No. 2004-161802 Patent Publication No. 7-324976 Patent Publication 2000-171291 Patent Publication 11-290323 Patent Publication 2005-33559 International Publication Number W001 / 070113

In conventional diagnostic devices and pattern recognition devices, the means for determining whether or not there is a measurement error in the measured signal is incomplete by checking only the presence or absence of signal overflow, or by visual inspection or intuition of the measurer. It is to determine the presence or absence of (measurement error). However, as described above, a diagnostic device or a pattern recognition device for processing a stationary signal is not sufficient to confirm whether or not there is an overflow of the measured signal, and it is necessary to check the continuity of the signal. . In addition, there is a limit to judging whether or not there is a defect by visual observation or intuition of the measurer, and in particular, in the case of a measuring device that cannot directly check the measured raw signal, the accuracy of the determination cannot be guaranteed.

The present invention relates to whether a measured signal is a stationary signal or a non-stationary signal before performing a process for diagnosis or pattern recognition in various diagnostic apparatuses and pattern recognition apparatuses in various fields. It is an object of the present invention to provide a signal inspection method and a signal inspection module for determining whether or not there is a measurement error in a signal that has been transmitted.

In order to solve the above problems, the present invention employs the following means. Whether the measured signal is a stationary signal or a non-stationary signal is inspected by confirming the steadiness of the dimensional feature parameter and the dimensionless feature parameter. In addition, whether or not there is a defect (measurement error) in the measured signal is inspected by confirming whether there is an overflow or outlier.

In the present invention, there is provided a method and a module for automatically inspecting whether a signal measured for state diagnosis or pattern recognition of an object is a stationary signal, a non-stationary signal, or whether there is a defect. Can be provided. Therefore, the method and module of the present invention can be applied to various diagnostic devices and pattern recognition devices in various fields, can improve the accuracy of state diagnosis and pattern recognition, and is useful for automation of signal inspection.

The processing flow of the present invention is shown in FIG . Here, the best mode of the present invention will be described along this flow.

Signal measurement (Fig. 1A)
Here, the target signals are a vibration signal, a voltage signal, a current signal, an acoustic signal, a biological signal, and the like, and are used for state diagnosis and pattern recognition of the target object.

Time series signal (B in Fig. 1)
The inspection target of the present invention is a time-series signal, and as shown in the examples of FIGS. 5, 6, 7, and 8, the target signal that is measured by the measurement unit and is intended to perform processing such as state diagnosis and pattern recognition This is the entire waveform data. That is, the stationarity of the signal is determined by checking whether or not the measured signal is defective, and whether or not the statistical properties of the time series signal change with time. The measured signal becomes N discrete data waveforms by sampling and is represented by x (i) (i = 1 to N). Incidentally, if the average value mu _r and the standard deviation S _r of x (i), by performing the normalization of x (i) as follows,

Overflow detection method (C and D in Fig. 1)
Overflow means that the maximum value (plus side) of the signal is larger than the maximum measurement range of the measurement device, or the minimum value (minus side) is smaller than the maximum measurement range of the measurement device. Whether an overflow has occurred is determined as follows. When the maximum value and the minimum value of the waveform data x (i) are x _max (i) and x _min (i), respectively, and the maximum measurement range and the minimum measurement range of the measuring device are R _max and R _min , respectively, x _max ( If i) ≧ R _{max or} x _min (i) ≦ R _min , it is determined that an overflow has occurred.

Outlier detection method (E, F, G, H in Fig. 1)
Outliers are defined as “extremely large values or extremely small values that deviate from the data population in statistics”. There are various outlier test methods. In the present invention, in order to speed up the processing, the waveform data has a value equal to or greater than “average value + 4 × standard deviation” or less than “average value−4 × standard deviation”. As the outlier is larger than “average value + 4 × standard deviation” or smaller than “average value−4 × standard deviation”, the degree of occurrence of the outlier is larger. There are two methods for detecting outliers.
(1) Outlier detection method using average value and standard deviation of waveform data
Let x (i) (i = 1 to N) be the discrete data of signals obtained for state diagnosis, pattern recognition, etc., and let | x (i) | be the absolute value of x (i). ) | of the average value and the respective standard deviations mu S, after the range of the coefficient k was set to _{k min ~k max, | x (} i) | value is in the | x (j) | (j = If 1 to N)> μ + k _min S, x (j) is determined as an outlier. In general, a _k min _~k max = _4~50. Desirable _k min _{~k max} is 5 to 20.

(2) Outlier detection method using absolute maximum value of waveform data
When the L absolute maximum values of x (i) are | x _Mj | (j = 1 to L ), and the average value and standard deviation after removing x _Mj are μ _a and S _a , respectively , the coefficient k after setting the range of _a and _{k amin ~k amax, | x Mj} | _{if> μ a + k min S a} , it is determined that values outside the _{x Mj.} Where in general _k amin _~k amax = _4~30. Desirable _k amin _{~k amax} is 5 to 20. In general, L / N is 0.00001 to 0.1, and a desirable L / N range is 0.0001 to 0.001.

(3) Calculation method of the degree of outlier occurrence
When the degree of occurrence of an outlier (hereinafter referred to as “outlier”) is expressed as a percentage, the upper limit value and lower limit value of the coefficient k (or k _a ) are set to k _max and k _min (or k _amax and k _amin ), respectively. When | x (j) | ≧ μ + k _max S (or | x _Mj | ≧ μ _a + k _amax S _a ), the degree of deviation is 100%, and | x (j) | ≦ μ + k _min S (or When | x _Mj | ≦ μ _a + k _amin S _a ), the degree of deviation is 0%, and | x (j) | = μ + k _min S to μ + k _max (or | x _Mj | = μ _a + k _amin S _a to In the case of μ _a + k _amax S _a ), if the degree of deviation is 0% to 100%, when the degree of deviation is larger than a preset threshold value, the signal is determined as an inadequate signal and appropriate measures are taken.
For _example, when _k max = 15 and _{k min} = _{5, 0%} if _{x max = μ + 5S, 50} % if _{x max = μ + 10S, 100} % if _{x max> μ} + 15S.
When the outlier generation degree is x% or more, it is determined that there is a signal defect or a measurement error, but x% is called “outlier percentage threshold” and is determined by the measurer according to the property of the measurement target. Further, after it has been found that the degree of occurrence of outliers is large, the measurer can decide whether to measure the signal again or to process the outliers as a countermeasure.

なお、上記の有次元特徴パラメータ以外に実効値もあるが、実効値の定常性については標準偏差と同様に考えることができる。On the continuity of dimensional feature parameters (I, J, K in Fig. 1)
The dimensional feature parameter is a parameter with a unit obtained from waveform data. Examples of dimensional feature parameters of the waveform data x (i) include the following.

Although there are effective values in addition to the dimensional feature parameters described above, the continuity of the effective values can be considered in the same way as the standard deviation.

いは、ｘ_ｊ（ｉ）とｘ_ｋ（ｉ））で表す。一般にＭ／Ｎ＝０．０００１＝０．５であり、望ましいＭ／Ｎの範囲は０．０１〜０．１である。第ｊ区間と第ｋ区間の波形データ数をＮ_ｊとＮ_ｋとする。

When checking the continuity of the dimensional feature parameter, the waveform data x (i) (or

_Or x _j (i) and x _k (i)). Generally, M / N = 0.0001 = 0.5, and a desirable M / N range is 0.01 to 0.1. The number of waveform data in the j-th section and the k-th section is N _j and N _k .

(1) Stationary evaluation method I for dimensional feature parameters I
When testing the continuity of μ, the average value in x _j (i) and x _k (i) is μ _j and μ _k, and the standard deviation in x _j (i) and x _k (i) is S _j and Assuming S _k , if μ _j = μ _k , the j-th section and the k-th section are defined as steady sections related to μ. If the average values μ in the M sections are all equal, that is, if μ ₁ = μ ₂ =... Μ _j = μ _k =... = Μ _M , the average value of the waveform data (signal) is stationary. Can be determined. However, in practice, taking into account the variation of the data waveform, giving significance level α probability, the hypothesis _{μ 1 = μ 2 = ··· μ} j = μ k = ··· = μ M statistical test for Do. Many statistical test methods [Non-Patent Document 1] have been proposed, but one example is shown here. For example, given a significance level α,

であるば、上記の仮説が成り立たなく、信号が非定常と判定される。ここで、ｊ≠ｋ，ｊ，ｋ＝１〜Ｍで、ｔ（α，∞）は自由度∞のｔ分布の確率密度関数が上側確率αに対するパーセント点である。αの最小値と最大値をそれぞれα _ｍｉｎ とα _ｍａｘ とすると、一般にα _ｍｉｎ 〜α _ｍａｘ ＝０．００００１〜０．９であるが、望ましい値としてα _ｍｉｎ 〜α _ｍａｘ ＝０．００１〜０．５である。

る。
Ｋ．Ａ．Ｂｒｏｗｎｌｅｅ：Ｓｔａｔｉｓｔｉｃａｌ Ｔｈｅｏｒｙ ａｎｄ Ｍｅｔｈｏｄｏｌｏｇｙ ｉｎ Ｓｃｉｅｎｃｅ ａｎｄ Ｅｎｇｉｎｅｅｒｉｎｇ，Ｓｅｃｏｎｄ Ｅｄｉｔｉｏｎ，Ｔｈｅ Ｕｎｉｖｅｒｓｉｔｙ ｏｆ Ｃｈｉｃａｇｏ，１９６５．

If this is the case, the above hypothesis does not hold and the signal is determined to be non-stationary. Here, j ≠ k, j, k = 1 to M, and t (α, ∞) is a percentage point with respect to the upper probability α of the probability density function of the t distribution with ∞ degrees of freedom. If α _min and α _max are α _min and α _max , respectively , α _{min to} α _max = 0.00001 to 0.9, but α _{min to} α _max = 0.001 to. 5.

The
K. A. Brownlee: Statistical Theory and Methodology in Science and Engineering, Second Edition, The University of Chicago, 1965.

(2) Steadyness evaluation method for dimensional feature parameters II

であるば、信号が非定常と判定される。ここで、ｔ（α／２，Ｎ_ｊ−１）は自由度（Ｎ_ｊ−１）のｔ分布の確率密度関数が上側確率α／２に対するパーセント点である。αの最小値と最大値をそれぞれα _ｍｉｎ とα _ｍａｘ とすると、一般にα _ｍｉｎ 〜α _ｍａｘ ＝０．００００１〜０．９であるが、望ましい値としてα _ｍｉｎ 〜α _ｍａｘ ＝０．００１〜０．５である。

If it is, it is determined that the signal is non-stationary. Here, t (α / 2, N _j −1) is a percentage point of the probability density function of the t distribution with the degree of freedom (N _j −1) with respect to the upper probability α / 2. If α _min and α _max are α _min and α _max , respectively , α _{min to} α _max = 0.00001 to 0.9, but α _{min to} α _max = 0.001 to. 5.

(3) Method for evaluating stationarity of dimensional feature parameters III
When testing the stationarity of S, if the standard deviations in x _j (i) and x _k (i) are S _j and S _k ,

であるば、Ｓ_ｊ≠Ｓ_ｋであり、信号が非定常と判定される。ここで、ここで、Ｆ（α／２，Ｎ_ｊ−１，Ｎ_ｋ−１）は自由度（Ｎ_ｊ−１，Ｎ_ｋ−１）のＦ分布の確率密度関数が上側確率α／２に対するパーセント点である。αの最小値と最大値をそれぞれα _ｍｉｎ とα _ｍａｘ とすると、一般にα _ｍｉｎ 〜α _ｍａｘ ＝０．００００１〜０．９であるが、望ましい値としてα _ｍｉｎ 〜α _ｍａｘ ＝０．００１〜０．５である。

る。

If so, S _j ≠ S _k and the signal is determined to be non-stationary. Here, F (α / 2, N _j −1, N _k −1) is a probability density function of F distribution with degrees of freedom (N _j −1, N _k −1) with respect to the upper probability α / 2. Percentage point. If α _min and α _max are α _min and α _max , respectively , α _{min to} α _max = 0.00001 to 0.9, but α _{min to} α _max = 0.001 to. 5.

The

と、Stationary evaluation method for dimensional feature parameters IV
(4) Stationary evaluation method for dimensional feature parameters IV

When,

であるば、信号が非定常と判定される。ここで、Ｘ^２（α／２，Ｎ_ｊ−１）は自由度（Ｎ_ｊ−１）のカイ２乗分布の確率密度関数が上側確率α／２に対するパーセント点である。αの最小値と最大値をそれぞれα _ｍｉｎ とα _ｍａｘ とすると、一般にα _ｍｉｎ 〜α _ｍａｘ ＝０．００００１〜０．９であるが、望ましい値としてα _ｍｉｎ 〜α _ｍａｘ ＝０．００１〜０．５である。

If it is, it is determined that the signal is non-stationary. Here, X ² (α / 2, N _j −1) is a percentage point of the probability density function of the chi-square distribution with the degree of freedom (N _j −1) with respect to the upper probability α / 2. If α _min and α _max are α _min and α _max , respectively , α _{min to} α _max = 0.00001 to 0.9, but α _{min to} α _max = 0.001 to. 5.

(5) Unsteady calculation method of dimensional feature parameters
When the degree of instability of a dimensional feature parameter (hereinafter referred to as “dimensional degree of instability”) is expressed as a percentage, when the significance level α _d is given, [Equation 2], [Equation 3], [Equation] 4] and [Equation 5] when the left side and the right side are equal, and α _d ≧ α _max , the dimensional instability is 0%, and when α _d ≦ α _min , the dimensional instability is about Is 100%, and when α _d = α _{min to} α _max , if the dimensional instability is 100% to 0%, the signal is an unstable signal when the dimensional instability is greater than a preset threshold. And take appropriate measures.
For example, in [Equation 2], α _max = 0.2 and α _min = 0.001. When α = 0.3, the left side and the right side of [Equation 2] are equal to 0%, and when α = 0.0005, the left side and the right side of [Equation 2] are equal to 100%, α = 0.1. When the left and right sides are equal, 50.3%, and when α = 0.05, the left and right sides are equal to 25.1%.
If the dimensional instability is equal to or greater than x%, if the signal is determined to be “unsteady”, x% is called the “non-stationary percentage threshold” of the dimensional feature parameter and is determined by the measurer according to the property of the measurement object. .

When it is determined that the signal is non-stationary, whether or not to re-measure the signal is determined by the nature of the object. In some cases, the result of the test is shown to the measurer, and the measurer decides whether to remeasure the signal.
It should be noted that it is obviously inappropriate to input a non-stationary signal to a diagnosis device or a pattern recognition device that can process only a stationary signal.

About the stationary evaluation method of dimensionless feature parameters (L, M, N in Fig. 1)
The dimensionless feature parameter is a unitless parameter obtained from waveform data. Examples of dimensionless feature parameters of the waveform data x (i) (= x _i ) include the following.

ここで、

here,

Here, mu is the average value of x _i, sigma is the standard deviation of x _i.

ここで，μ_ｐは波形の極大値（ピーク値）の平均値である。

Here, μ _p is an average value of the maximum value (peak value) of the waveform.

ここで、μ_ｍａｘはの１０個最大値の平均値である。

Here, μ _max is an average value of 10 maximum values.

ここで、σ_ｐは極大値の標準偏差値である。

Here, σ _p is the standard deviation value of the maximum value.

ここで、μ_Ｌとσ_Ｌはそれぞれ極小値（谷値）の平均値と標準偏差値である。

Here, μ _L and σ _L are an average value and a standard deviation value of local minimum values (valley values), respectively.

但し、ｘ_ｋｉ＞ｋσ（例えば、ｋ＝１，２）。

However, x _ki > kσ (for example, k = 1, 2).

但し、ｘ_ｈｉ＞ｋσ。（例えば、Ｈ＝１，２）。
なお、以下で述べる無次元特徴パラメータの定常性の判定法は上記以外に他の無次元特徴パラメータにも適用できる。

However, x _hi > kσ. (For example, H = 1, 2).
The method for determining the continuity of a dimensionless feature parameter described below can be applied to other dimensionless feature parameters in addition to the above.

When examining the continuity of the dimensionless feature parameter, the waveform data x (i) is divided into M sections, and the waveform data of the jth section is represented by x _j (i). Generally, M / N = 0.0001 to 0.1, and a desirable M / N range is 0.005 to 0.01. Let N _j be the number of waveform data in the jth section. The dance dimension feature parameter is uniformly represented by p, and the dimensionless feature parameter p in x _j (i) is represented by p _j (j = 1 to M).
When determining the stationarity of the dimensionless feature parameter p, there are the following methods.

(1) Stationarity evaluation method for dimensionless feature parameters I
After the range of the advance coefficient _{k p} was set to _{k pnin ~k pmax, | p j} |> μ p + k pmin S p is determined if _{p j} and a specific dimensionless characteristic parameter signal is determined to unsteady The Generally a _{k pnin ~k} pmax = _1~20. Desirable _{k pnin ~k pmax} is 2-3.

(2) Method for evaluating stationarity of dimensionless feature parameters II
After the range of the advance coefficient _{k L} was set to _k Lnin _{~k Lmax,} | is _{determined> k Lmin} if dimensionless characteristic parameter p unstable and | _{S p} / mu _p. In general, a _k Lnin _~k Lmax = _0.1~6. Desirable _k Lnin _{~k Lmax} is 0.5 to 3.

(3) Method for evaluating stationarity of dimensionless feature parameters III
The dimensionless feature parameter in x _j (i) is set to p _Mj (i = 1 to U) as the U maximum value of p _j , and the average value and standard deviation of x (i) after removing p _Mj are μ When _pn and _{S pn,} after the range of the advance coefficient _{k M} was set to _{k Mnin ~k Mmax, | p Mj} | determined> the mu _pn + _{k Mmin} S _pn If dimensionless characteristic parameter p unstable and, signal Is determined to be non-stationary. In general, k _Mmin -k _Mmax = 2-10. Desirable _k Mnin _{~k Mmax} is 3-6. In general, U / M = 0.001 to 0.5, and a desirable U / M range is 0.001 to 0.01.

(4) Method for evaluating the continuity of dimensionless feature parameters IV
After the range of the advance coefficient _{k i} was set to _{k inin ~k imax, | p j} × p j-1 | (j = 2~M) the maximum _{value> k imin} If dimensionless characteristic parameter p of the free and Determined. how to determine the k _i experience

(5) Non-stationary calculation method for dimensionless feature parameters
When expressing the non-steady degree of the dimensionless feature parameter (hereinafter referred to as “the dimensionless instability degree”) as a percentage, the upper limit value and the lower limit value of the coefficient k _p (k _L , k _M , and k _i are also the same) When k _max and _kmin, and | p _j | ≧ μ _p + k _pmax S _p , the dimensionless _instability degree is 100%, and when | p _j | ≦ μ _p + k _pmin S _p , dimensionless _instability the extent to 0%, when _{k p = μ p + k pmin} S p ~k pmax S p, when the dimensionless unstable about to 0% to 100%, when the threshold is larger than the dimensionless astable about previously set The signal is determined to be an unstable signal and appropriate measures are taken.
For _example, when _{k pmax} = 5 and _{k pmin = 1, | p j} | if _{= μ p + 0.8S p 0%} , | p j | = if _{u p + 3S p 50%,} | p j | = 100% if it is μ _p + 6S _p.
Further, the upper limit value and lower limit value of the coefficient k _i (the following method can also be applied to k _L ) are set as k _imax and k _imin , respectively, and | p _j p _j−1 | (j = 2 to M). when the maximum value is referred to as Delta] p _jmax, when Δp _jmax ≧ _{k imax,} to 100%, when Δp _{jmax <k imin,} and 0%.
For example, when k _imax = 11 and k _imax = 1, 0% if Δp _jmax = 0.9, 50% if Δp _jmax = 6, and 100% if Δp _jmax = 12.
If the non-steady degree of the dimensionless feature parameter is greater than or equal to x%, if the signal is determined to be “unsteady”, the x% is called the “non-stationary percentage threshold” of the dimensionless feature parameter and is measured according to the property of the measurement object. Decide.

When it is determined that the signal is non-stationary, whether or not to re-measure the signal is determined by the nature of the object. In some cases, the result of the test is shown to the measurer, and the measurer decides whether to remeasure the signal.
It should be noted that it is obviously inappropriate to input a non-stationary signal to a diagnosis device or a pattern recognition device that can process only a stationary signal.

As described above, the result of the signal inspection by each item is shown to the signal measurer, and it can be determined by the measurer whether or not the signal is measured again. When evaluating with a plurality of existence / non-dimension feature parameters, the maximum value of the obtained non-stationary degree (%) of the existence / non-dimension feature parameters is displayed as a final result.
FIG. 2 shows a method (I, II) of displaying the signal inspection result by a bar graph.
FIG. 3 shows a display method (I, II, III, IV) of the signal inspection result by the lamp. Whether the lamp is turned on or off is determined by the characteristics of the object. A desirable way to decide is as follows.
In the display methods I and III of the signal inspection result by the lamp, 80% or more is “red”, 30% to 80% is “yellow”, and less than 30% is “blue”. In the display methods II and IV of the signal inspection result by the lamp, 50% or more is “lighted” and less than 50% is “lighted out”.

In addition, a buzzer sound can be displayed. If the signal inspection result (%) exceeds the set percentage, the buzzer will sound. For example, 50% or more is a “buzzer sound”.

FIG. 4 shows a block diagram of a signal inspection module constructed by the method of the present invention. That is, a signal indicating whether the signal measured by the signal measurement unit is a stationary signal, a non-stationary signal, or whether there is a measurement error in the measured signal for the purpose of state diagnosis or pattern recognition of the object. Automatically inspected by the inspection unit, the inspection result is displayed on the display unit of the inspection result, and the signal measurer based on the inspection result before performing the diagnosis or pattern recognition processing by the diagnosis unit or the pattern recognition unit Whether the signal is measured again can be determined on the spot, or can be automatically handled by the signal inspection unit according to a rule set in advance, and accuracy such as state diagnosis and pattern recognition can be ensured.

FIG. 5 is a vibration acceleration signal of the rotating machine measured by the equipment diagnosis apparatus. This equipment diagnosis device is used for state diagnosis of rotating machines such as a motor, a pump, a generator, and a gear device of steady rotation. This diagnostic apparatus measures a steady vibration signal of a rotating machine to be diagnosed and performs a state diagnosis by processing the steady vibration signal.
The results of examining this signal are as follows.
Data measurement conditions: number of data: N = 2048, sampling frequency: 1000 Hz.
Overflow inspection result: Overflow detection ranges are R _max = 4.0 (v) and R _min = −4.0 (v). In this signal, x _max (i) = 2.79 (v) and x _min (i) = − 2.95 (v). Therefore, there is no overflow from x _max (i) ≦ R _max and x _min (i) ≧ R _min .
Outlier test result: The coefficient k of the outlier detection method I is set to _kmax = 20 and _kmin = 5. Μ of this signal is 0.013784 and S: 0.799. Therefore, since the maximum value of | x (i) | 2.79 <0.0137784 + 5 × 0.799, the degree of outlier is 0%.
Yes dimensional feature parameters (mu _r and mu _a) constancy of the determination result of:
Waveform data division number M = 8,
Significance level α = 0.001 in the dimensionality feature parameter stationaryity evaluation method I,
Minimum value of significance level α _min = 0.001,
Maximum value of significance level α _min = 0.01,
t (0.001,2048 / 8) = 3.329867,
t (0.01,2048 / 8) = 2.596,
Mean mu ₁ and _x 1 (i) of the first divided section _| x 1 (i) _|, based on the average value mu _a1 of, _{t i} of the divided sections obtained by routine evaluation method I chromatic dimensional feature parameters And the determination result by t _ai is as follows.
T ₂ = 1.45 of second divided section, unsteady degree = 0%
T _{a2 of} the second divided section = 0.80, unsteady degree = 0%
T ₃ = 0.23 of the third divided section, unsteady degree = 0%
T _{a3 in} the third divided section = 0.97, unsteady degree = 0%
T ₄ = 1.35 of the fourth divided section, unsteady degree = 0%
T _{a4 in} the fourth divided section = 0.87, unsteady degree = 0%
T ₅ = 0.79 of the fifth divided section, unsteady degree = 0%
T _{a5 of} the fifth divided section = 0.10, unsteady degree = 0%
T ₆ = 2.36 of the sixth divided section, unsteady degree = 0%
T _a6 = 1.08 in the sixth divided section, unsteady degree = 0%
T ₇ = 0.40 of the seventh divided section, unsteady degree = 0%
T _{a7 in} the seventh divided section = 0.25, unsteady degree = 0%
T ₈ = 1.95 in the eighth divided section, unsteady degree = 0%
T _a8 = 1.07 in the eighth divided section, unsteady degree = 0%
Based on the above results, this waveform has a non-stationary degree of dimensional feature parameters of 0%.
Determination result of stationarity of dimensionless feature parameters (p ₂ and p ₃ ):
Waveform data division number M = 8,
_{K j} at steady Evaluation Method IV dimensionless characteristic parameter coefficient _{p 2 k 2max = 2, k} 2min = 1
The coefficient of p ₃ k _3max = 3, k _3min = 1
Δp _{jmax of} p ₂ = 0.89, unsteady degree = 0%
Δp _{jmax of} p ₃ = 1.23, unsteady degree = 11.5%
Based on the above results, this waveform has a non-stationary degree of dimensionless feature parameter of 11.5%.
Therefore, this signal was used for condition diagnosis as a normally measured signal.

6 is an example of a waveform measured in the same diagnostic apparatus and the measurement conditions and the waveform of FIG. The results of examining this signal are as follows.
Data measurement conditions: number of data: N = 2048, sampling frequency: 1000 Hz.
Overflow inspection result: Overflow detection ranges are R _max = 4.0 (v) and R _min = −4.0 (v). In this signal, x _max (i) = 3.14 (v) and x _min (i) = − 3.19 (v). Therefore, there is no overflow from x _max (i) ≦ R _max and x _min (i) ≧ R _min .
Outlier test result: The coefficient k of the outlier detection method I is set to _kmax = 20 and _kmin = 5. This signal has μ = 0.26 and S = 0.701. Therefore, from the maximum value of | x (i) | 3.19 <0.26 + 5 × 0.701, the degree of outlier is 0%.
Yes dimensional feature parameters (mu _r and mu _a) constancy of the determination result of:
Waveform data division number M = 8,
Significance level α = 0.001 in the dimensionality feature parameter stationaryity evaluation method I,
Minimum value of significance level α _min = 0.001,
Maximum value of significance level α _min = 0.01,
t (0.001,2048 / 8) = 3.329867,
t (0.01,2048 / 8) = 2.596,
Mean mu ₁ and _x 1 (i) of the first divided section _| x 1 (i) _|, based on the average value mu _a1 of, _{t i} of the divided sections obtained by routine evaluation method I chromatic dimensional feature parameters And the determination result by t _ai is as follows.
T _{2 of} the second divided section = 4.98, unsteady degree = 100%
T _{a2 of} the second divided section = 0.78, unsteady degree = 0%
T _{3 of} the third divided section = 5.32, unsteady degree = 100%
T _a3 = 1.91 of the third divided section, unsteady degree = 0%
T ₄ = 4.48 of the fourth divided section, unsteady degree = 100%
T _{a4 in} the fourth divided section = 0.82, unsteady degree = 0%
T ₅ = 4.92 in the fifth divided section, unsteady degree = 100%
T _{a5 of} the fifth divided section = 0.10, unsteady degree = 0%
T ₆ = 0.44 in the sixth division, unsteady degree = 0%
T _{a6 of} the sixth divided section = 0.76, unsteady degree = 0%
T ₇ = 2.75 of the seventh divided section, unsteady degree = 5%
T _{a7 in} the seventh divided section = 0.33, unsteady degree = 0%
T ₈ = 6.72 in the eighth divided section, unsteady degree = 100%
T _{a8 of} the eighth divided section = 0.99, unsteady degree = 0%
Based on the above results, this waveform has a non-stationary degree of the dimensional feature parameter of 100%.
Determination result of stationarity of dimensionless feature parameters (p ₂ and p ₃ ):
Waveform data division number M = 8,
_{K j} at steady Evaluation Method IV dimensionless characteristic parameter coefficient _{p 2 k 2max = 2, k} 2min = 1
The coefficient of p ₃ k _3max = 3, k _3min = 1
Δp _{jmax of} p ₂ = 1.68, unsteady degree = 68%
Δp _{jmax of} p ₃ = 1.75, unsteady degree = 37.5%
Based on the above results, this waveform has a non-stationary degree of dimensional feature parameters of 68%.
In this example, since the degree of unsteady dimensional feature parameter is 100%, it is determined that the equipment diagnosis apparatus that can process only a steady signal cannot process this signal.

Figure 7 is also a waveform example measured in the same diagnostic apparatus and the measurement conditions and the waveform of FIG. The results of examining this signal are as follows.
Data measurement conditions: number of data: N = 2048, sampling frequency: 1000 Hz.
Overflow inspection result: Overflow detection ranges are R _max = 2.0 (v) and R _min = −2.0 (v). In this signal, x _max (i) = 3.17 (v) and x _min (i) = − 2.79 (v). Therefore, from x _max (i)> R _max and x _min (i) <R _min , it was determined that an overflow occurred, and the signal was measured again.

Figure 8 is also an example of a waveform measured in the same diagnostic apparatus and the measurement conditions and the waveform of FIG. The results of examining this signal are as follows.
Data measurement conditions: number of data: N = 2048, sampling frequency: 1000 Hz.
Overflow inspection result: Overflow detection ranges are R _max = 15.0 (v) and R _min = -15.0 (v). In this signal, x _max (i) = 9.7 (v) and x _min (i) = − 14.0 (v). Therefore, there is no overflow from x _max (i) ≦ R _max and x _min (i) ≧ R _min .
Outlier inspection result: L in the outlier detection method II is L = 2, and the coefficient k is k _max = 12 and _kmin = 6. Μ of the signal is −0.0054 and S = 0.822. Therefore, from the maximum value 12.3> −0.0054 + 12 × 0.822 of | x (i) |, the degree of the outlier is 100%, and the signal was measured again.

It is a graph which shows the flow of processing of the present invention. It is a graph which shows the display method of the signal inspection result by a bar graph. It is a graph which shows the display method of the signal test | inspection result by a lamp | ramp. It is a graph which shows the structure of the signal inspection module of this invention. It is a graph which shows the example of a steady signal. It is a graph which shows the example of an unsteady signal. It is a graph which shows the example of the signal which overflowed. It is a graph which shows the example of the signal in which an outlier exists.

Regarding the reference numerals in FIG.
DESCRIPTION OF SYMBOLS 1 State diagnostic apparatus and pattern recognition apparatus, 2 Signal inspection module of this invention, 3 Signal measurement part, 4 Signal inspection part, 5 Inspection result display part, 6 Diagnosis part and pattern recognition part

Claims (3)

Perform comprehensive evaluation of the waveform data of signals measured from the object for condition diagnosis and pattern recognition using the presence of overflow and the following “outlier value calculation method” and “waveform instability calculation method” below. To determine whether the waveform data can be used for state diagnosis or pattern recognition.
"Outlier calculation method":
The waveform data of signals obtained for state diagnosis and pattern recognition is x (i) (i = 1 to N), the absolute value of x (i) is | x (i) |, and | x (i) | average value and the standard deviation of the respectively μ and S, then the range of the coefficient k was set to _{k min ~k max, | x (} i) | value is in the | x (j) | (j = 1 ˜N)> μ + k _min S, x (j) is determined as an outlier,
Alternatively, if the L absolute maximum values of x (i) are | x _Mj | (j = 1 to L), and the average value and the standard deviation after excluding x _Mj are μ _a and S _a , respectively, the coefficient k after setting the range of _a and _{k amin ~k amax, | x Mj} |> if μ _a + _{k amin} S _a, determines that values outside the _{x Mj,}
When the degree of occurrence of an outlier (hereinafter referred to as “outlier”) is expressed as a percentage, the upper limit value and lower limit value of the coefficient k (or k _a ) are set to k _max and k _min (or k _amax and k _amin ), respectively. When | x (j) | ≧ μ + k _max S (or | x _Mj | ≧ μ _a + k _amax S _a ), the degree of deviation is 100%, and | x (j) | ≦ μ + k _min S (or When | x _Mj | ≦ μ _a + k _amin S _a ), the degree of deviation is 0%, and | x (j) | = μ + k _min S to μ + k _max S (or | x _Mj | = μ _a + k _amin S _a when ~μ _a + _{k amax} S _a) determining, when the degree of off-set to 0% to 100%, and deficiencies signal the signal when the threshold is greater than the order of disconnected preset.
"Calculation method of waveform instability":
“The method of calculating the degree of waveform instability” includes the following “(1) Method of calculating the degree of waveform instability based on the stability of dimensioned feature parameters” and “(2) Calculation of the degree of waveform instability based on the stability of dimensionless feature parameters” The signal obtained for condition diagnosis, pattern recognition, etc. was calculated using the following “(1) Method for calculating the degree of waveform instability based on the stability of dimensional feature parameters”. When the non-steady degree of the dimensional feature parameter is larger than a preset threshold value, or the dimensionless feature parameter calculated using the following “(2) Method for calculating the degree of waveform instability by the stability of the dimensionless feature parameter” If the unsteady degree is greater than a preset threshold, the signal is determined to be an unstable signal.
(1) Method for calculating the degree of waveform instability based on the stability of dimensional feature parameters:
The waveform data x (i) of a signal obtained for state diagnosis or pattern recognition is divided into M sections, and the waveform data of the jth section and the kth section are respectively x _j (i) and x _k (i). The number of waveform data in the j-th section and the k-th section is N _j and N _k , respectively, and the average values of dimensional feature parameters in the j-th section and the k-th section are μ _j and μ _k , respectively. And the standard deviations of the dimensional feature parameters in the k-th interval are S _j and S _k , respectively, and the average values μ _i (i = ₁ to M) in the M intervals are all equal (that is, μ ₁ = μ ₂ ). Μ _j = μ _k =... = Μ _M ) or standard deviations S _i (i = ₁ to _M ) in M intervals are all equal (ie, S ₁ = S ₂ =... S _j = S _k = ... = given the null hypothesis that S _M) in order to test the statistical theory, et al. After the maximum and minimum values of the significance level alpha set the alpha _max and alpha _min respectively, the alpha at the time when the null hypothesis is just discarded when changing the alpha from 0 to 1 alpha _d Then, when α _d <α _min , the dimensional feature parameter is determined to be non-stationary,
When the degree of instability of a dimensional feature parameter (hereinafter referred to as “the degree of dimensional instability”) is expressed as a percentage, when α _d ≦ α _min , the degree of dimensional instability is set to 100%, and α _d ≧ α _max When the dimensional instability is 0%, and when α _d = α _{min to} α _max , the dimensional instability is 100% to 0%, the dimensional instability is greater than a preset threshold. Sometimes the signal is determined to be an unstable signal.
(2) Waveform instability calculation method based on the stability of dimensionless feature parameters:
The dimensionless feature parameter calculated from the waveform data x (i) of the signal obtained for state diagnosis or pattern recognition is defined as p, and the average value and standard deviation of the dimensionless feature parameter p in x (i) are respectively represented. and mu _p and _{S p,} with x (i) was divided into M sections, the number of waveform data of the j-section and _{N j,} the waveform data of the j interval _{x j (i) (j =} 1~M) And the dimensionless feature parameter in x _j (i) is represented by p _j ,
“Waveform instability calculation method based on stability of dimensionless feature parameter” is the following “Dimensionless nonstationary degree calculation method 1”, “Dimensionless nonstationary degree calculation method 2” below, and “Dimensionless nonstationary degree calculation method” below. Degree calculation method 3 ”and“ Dimensionless non-stationary degree calculation method 4 ”described below. The following“ Dimensionless non-stationary degree calculation method 1 ”and“ Dimensionless non-stationary degree calculation ”described below are applied to the signal. The non-stable degree of the dimensionless feature parameter calculated using at least one of “Method 2”, “Dimensionless non-stationary degree calculation method 3” and “Dimensionless non-stationary degree calculation method 4” described below is determined in advance. If it is greater than the set threshold, the signal is determined to be an unstable signal.
Dimensionless unsteady degree calculation method 1:
After the range of the coefficient k _p is set as k _pnin to k _{pmax in} advance, if | p _j |> μ _p + k _pmin S _p , p _j is determined as a unique dimensionless feature parameter,
Unregulated about dimensionless characteristic parameter (hereinafter, referred to as "non-dimensional unstable about") When expressed as a percentage of, | _{p j} | when _{≧ μ p + k pmax S p} , the dimensionless unstable about 100% , | P _j | ≦ μ _p + k _pmin S _p , the dimensionless instability degree is set to 0%, and when | p _j | = μ _p + k _pmin S _{p to} μ _p + k _pmax S _p , dimensionless instability If the degree is 0% to 100%, the signal is determined to be an unstable signal when the dimensionless astable degree is larger than a preset threshold value.
Dimensionless unsteady degree calculation method 2:
After setting the range of the coefficient k _L as k _Lnin to k _{Lmax in} advance, if | S _p / μ _p |> k _Lmin , the dimensionless feature parameter p is determined to be unstable,
When the non-stable degree of the dimensionless feature parameter (hereinafter referred to as “the dimensionless non-stable degree”) is expressed as a percentage, when | S _p / μ _p | ≧ k _Lmax , the dimensionless non-stable degree is defined as 100%. When | S _p / μ _p | ≦ k _Lmin , the dimensionless instability degree is 0%, and when | S _p / μ _p | = k _{Lmin to} k _Lmax , the dimensionless instability degree is 0% to 100%. Then, when the dimensionless instability is larger than a preset threshold value, the signal is determined to be an unstable signal.
Dimensionless unsteady degree calculation method 3:
The dimensionless characteristic parameter of x _j (i) the number U of the maximum value of _{p j} and _p Mj (j = _1~U), respectively the mean and the standard deviation of x (i) after removal _{p Mj} mu When _pn and _{S pn,} in advance after the range of the coefficient _{k M} was set to _{k Mnin ~k Mmax, | p Mj} |> a mu _pn + _{k Mmin} S _pn If dimensionless characteristic parameter p is determined that unstable,
When expressing the non-stable degree of the dimensionless feature parameter (hereinafter referred to as “the non-dimensional non-stable degree”) as a percentage, when | p _Mj | ≧ μ _pn + k _Max S _pn , the non-dimensional non-stable degree is defined as 100%. , | P _Mj | ≦ μ _pn + k _Min S _pn , the dimensionless instability degree is set to 0%, and when | p _Mj | = μ _pn + k _Mmin S _{pn to} μ _pn + k _Max S _pn , dimensionless instability If the degree is 0% to 100%, the signal is determined to be an unstable signal when the dimensionless astable degree is larger than a preset threshold value.
Dimensionless unsteady degree calculation method 4:
After setting the range of the coefficient k _i as k _inin to k _{imax in advance,} if the maximum value of | p _j × p _j−1 | (j = 2 to M)> _kimin , the dimensionless feature parameter p is assumed to be unstable. Judgment,
When the non-stable degree of the dimensionless feature parameter (hereinafter referred to as “the dimensionless non-stable degree”) is expressed as a percentage, the maximum value of | p _j × p _j−1 | (j = 2 to M) ≧ k _imax When the dimensionless instability degree is 100% and the maximum value of | p _j × p _j−1 | (j = 2 to M) ≦ _kimin , the dimensionless _instability degree is 0%, and | p _j × _p j-1 | when the maximum value ₌ k _{imin to k imax} of (j = 2 to M), when the dimensionless unstable about to 0% to 100%, than a threshold dimensionless astable about previously set When the signal is large, the signal is determined as an unstable signal. 2. A first step of obtaining a waveform data by measuring a signal at a set sampling time by a signal measuring unit for state diagnosis and pattern recognition, and a second step of performing a signal inspection using the signal inspection module according to claim 1. A third step of displaying a result of the signal inspection on the display unit, a fourth step of determining whether or not the waveform data is used for state diagnosis or pattern recognition based on the result of the signal inspection, and the final step And a fifth step of starting signal remeasurement or state diagnosis, pattern recognition, and the like according to the determination result. An apparatus for state diagnosis and pattern recognition, comprising a signal inspection function based on the signal inspection method according to claim 2.

Images