JPS59142424A - Abnormality detecting device - Google Patents

Abstract

PURPOSE:To detect abnormality of a measuring signal without delay by forecasting a low-frequency noise and using it as a criterion, by using an adaptive type filter formed by combining Haar conversion and reverse Haar conversion, a self- recursive model, etc. CONSTITUTION:An A/D-converted measuring signal from a data buffer 3 is processed by an adaptive type filter in which a high-speed Haar converter 4, a high-speed reverse Haar converter 6, etc. are combined and installed through an error calculator 7, a spectrum selector 5, etc., a high frequency noise in the measuring signal is eliminated, and also a pulse signal for showing the abnormality becomes an emphasized signal. This signal is forecast and processed by a self-recursive model of a forecasting calculator 8, a low-frequency noise of the measuring signal is forecast, and a deciding device 9 decides an output of the adaptive type filter by basing on this forecasting value as a reference. If the forecasting fails, generation of abnormality which is not a noise in the measuring signal is detected, and the abnormality of the measuring signal can be detected without delay.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an abnormality detection device that detects abnormal or unique signals that suddenly appear in measurement signals based on measurement signals obtained from various sensors.

The following is a description of an example in which an abnormality in a power generation device is detected from a signal from a vibration accelerometer attached to a power generation device or the like.

Conventional abnormality detection devices that use measurement signals from vibration accelerometers mainly calculate the distribution of the power spectrum of the measurement signals.
It was based on the principle of detecting the presence of abnormal frequency spectra.

Normally, in order to obtain the distribution of this power spectrum, an analog resonant circuit is provided, and the power of the resonant voltage is measured by sweeping the resonant frequency.

This method determines whether an abnormality has occurred in the device, and
K was well suited to find the frequency of vibration associated with the anomaly. However, since abnormalities are detected by frequency spectrum analysis of measurement signals, detection is performed after the abnormality occurs.

This was insufficient for the purpose of detecting the occurrence of an abnormality without delay.

In order to eliminate these drawbacks, this invention combines Baar transform and inverse Baar transform to construct an adaptive low-pass filter whose pass band automatically changes according to the characteristics of the measurement signal, thereby smoothing the measurement signal. , calculate the sample variance by regarding the difference between this smoothed value and the measured signal as noise with a mean value of 0, and further predict the change in the smoothed value by predicting and calculating the low-frequency changes in the smoothed value using an autoregressive model, By comparing this predicted value with the smoothed value and the sample variance value, it is attempted to detect without delay the occurrence of an abnormal signal that is not noise in the measurement signal.

The present invention will be explained below with reference to examples thereof.

FIG. 1 shows an example of a measurement signal from a vibration accelerometer attached to a power generation device. In the figure, X is a measurement signal of a vibration accelerometer, and a pulse-like waveform surrounded by a circle (1) is a signal indicating the occurrence of an abnormality. The abnormality detection device according to the present invention attempts to detect the pulse-like signal (1) in the measurement signal X without delay.

FIG. 2 shows an embodiment of the abnormality detection device according to the present invention. (2) in the figure is the measurement signal X of the vibration accelerometer
This is a ~Φ converter that converts the signal into a digital signal every cycle Ts, and XK indicates the value of the measurement signal X at time KT11.

(3) is a rotator huffer that stores the digital signal n while updating it in the past from the present time to the time (N-t)Ts every period Ts;
The current value of is ψN-1, and the past value of 'l'8 is ψN-
1. In this way, set the past value to 90 by (N-t)Ts.

Output these.

(4) is a high-speed Var converter, which outputs ψ0.0 from the data buffer (3). ψ1. ..., ψN-1 is transformed into a Varl spectrum Co, Cm, ..., CN-
Output t. However, N is a positive number given in units of 2.

N=2” 1ll (5
) is based on the order M of the Baar spectrum specified by the 1-dispersion calculator (7) in the spectrum selector. C1,..., CM-1
Select and output.

(6) is a highway Bahr transformer which performs inverse Bahr transformation on the subsequences Co, Cs, . . . , CM-1 of the Baar spectrum, and outputs the output ψ0. ψ
1. ....

ψN-1 is approximated by M step-like functions, and approximate digital signals FO, Fl, . . . , FM-1 are output.

(7) is an error calculator that specifies the order M of the subsequence of the Var spectrum to be selected in the spectrum selector (5).

At this time, approximate digital signals Fo, F1 . ..., the digital signal ψ0. which is the output of the data buffer (3) of FM-t. ψ1
．． ....

Calculate the approximation error for ψN-1. Error calculator (7
) sequentially increases the order of the subsequence of the Baar spectrum specified to the spectrum selector (5) from M==1, calculates the approximation error each time, and determines whether the approximation error is smaller than the predetermined optimal error C, The degree M stops increasing when it becomes closest to e. If the order at this time is wealth, then the approximate digital signal Fo.

Fl, . . . , 6-1 are the digital signals ψ0. ψ1
．． ....

ψN-1 is optimally approximated, and the current value rr-t of this approximate digital signal is output as a smoothed value YK, and the approximation error is output as a variance σ2.

In this way, the high-speed var converter (4), the spectrum selector (5), the highway var converter (6), and the error calculator (7) are activated every time the stored value of data buffer 1 (2) is updated. The digital signal ψ0. Approximate digital signals Fo, F1 . . . , which optimally approximate ψl, . ...
, +1, and output the smoothed value YX =IIFX- and the variance σ2.

Change the passband according to the characteristics of the above measurement signal X.

Construct an adaptive filter that removes high-frequency noise.

(8) is a prediction calculator, which is calculated by the expressway bar converter (6) every time the stored value of the data buffer (3) is updated.

The output smoothed value YK is stored, and the past values are used to perform predictive calculation using an autoregressive model, and a predicted value fK of the smoothed value YK is output. This predicted value fK is a value close to the smoothed value YK when the measurement signal At that point, the autoregressive model can no longer predict the value, and it becomes significantly different from the smoothed value YK.

(9) is a discriminator, which compares the smoothed value YK and the predicted value fK, and when the difference becomes larger than the threshold value set using the variance σ2, a pulse indicating an abnormality in the measured signal A detection signal dK is output to notify that the signal (1) has occurred.

In this manner, the abnormality detection device shown in FIG. 2 detects without delay the occurrence of an abnormal signal that is not noise in the measurement signal of the vibration accelerometer.

The operation of each block in FIG. 2 will be explained in detail below using mathematical formulas and diagrams.

FIG. 3 is a diagram illustrating the principle of approximation and smoothing of measurement signals using step-like functions. First, the data buffer (3
) are regarded as a sequence of N numbers.

ψ0. ψ1. ...-1ψN-1 (2 plus ψ0=XK-N+1.911=XK-N+2.
...-, 9) N-1=store.

Since the sequence of equation (2) forms a step-like function with time as the horizontal axis, M step-like functions that approximate this are applied using the method of least squares, and the sequence is yo, Ft+..., FM-113). In this way, the error between the two is (41, +51 formula% formula% (5) where N==20.M==?, m and n are positive integers and n≧m.

Mathematically, the inequality of equation (6) clearly holds true.

ρ0≧ρl≧...≧ρm≧...≧ρn (6) Then, if a positive number 6 is given as the optimal error, there exists the smallest m such that ρm≦I, so this is made exponentially dense.

=2" +71, the sequence FO, Fl, ..., FZ-t
(81 is the condition of M=-2", the error ρm is less than C,
Furthermore, the approximate digital signal "+Fl, . . . , r" which becomes the step-like function closest to ε and optimally approximates the value of the digital signal
i-1 is obtained.

These approximate digital signals FO, Fl, ..., nX-
The current value FW-t of t is set as the smoothed value Yx, and the approximate digital signals Fo, F1 .
..., if FEr-t is calculated and the smoothing value is updated, the smoothing value... will be an output that greatly smoothes the change in the digital signal xK when the optimum number of steps is small, but if 1M becomes large, the degree of smoothing will change becomes smaller, and when 1=N, it becomes the digital signal XK itself. i.e. data buffer (3)
The output ψ0. ψ1. ..., when there is no pulse-like signal (1) indicating an abnormality in ψN-1, the optimal number of steps i becomes small, greatly smoothing out the noise, but when the pulse-like signal (1) is included, the optimal number of steps i becomes As the smoothing effect becomes larger, the pulsed signal (1) is output as is.

In other words, the digital signal is passed through a type of adaptive filter whose passband changes according to changes in its characteristics, achieving effective filter processing with little delay.

The digital signal ψ0. ψ1. ..., 91N-1
The adaptive filter approximating by a step-like function is composed of a fast Varl transformer (4), a spectrum selector (5), a fast Varl transformer (6), and an error calculator (7).

First, the high-speed Var converter (4) converts the digital signal ψ0. ψ1. River, ψN−1
w/%-ru conversion t/'-rusuhekl-ru co,
Output Ct, ..., CN-1.

Here, the Barr function χ is a function defined as follows in the interval (0, 1).

Let t be a variable in the interval (0, 1).

χo(t)=1. 0≦t≦1 (9) FIG. 4 depicts a part of the Barr function of the equation (9).

Using the Barr function, the above digital signal ψ0゜ψ1
．． ..., ψN-1 is known to be expressed by equation 01.

(j=0.1...,, N-1; N=2°) where C
o, cP are nor)-nor spectra, and the relationship with the above bar spectra Co, Cs,..., CN-t +
-It will look like 1ua type.

Furthermore, using equation 01), the digital signal ψ0. ψl
.

..., from ψN-1 to Baar spectrum Co, C7(
[Use the following fast Var transformation to find o. First, the digital signal ψ0. ψ1. ..., ψN-1 is regarded as a sequence and the following recurrence formula is calculated.

ψi(0'-9'J, (j=0.1.-・-, N
-1) G3)'PJ"-ψ2r-"+'P
::+”t), (p=1.2. ・=, n-1:
q=0.1. −.

z"-!Lx) (14) Then, using the obtained ψj(0)・ψ, +p+, 2'C
o-9'0(n-1)+ ψj"-1) (15) 2G mouth c(・+1)=ψJ:'"1>-92's'+1"n-
r+1Q61 (r=1.21...+n; B=Q, i,
HHH, 21121) (Sequence C by formula 151G61
,(2"-1), ,・, CP C2", Ct"'
, c.

By sequentially calculating the Baar spectrum Co.

CP, that is, Co, Cs, . . . CN-t is determined.

The high-speed bar converter (4) shown in FIG.

Next, the spectrum selector (5) is a high-speed bar converter (4).
) Baar spectra obtained with Co, C1,...,
Based on the order M specified by the error calculator (7) from CN-t, subsequences Co, Cs, . . . of the Baar spectrum are calculated.
Select and output CM-1.

Next, the expressway Bahr converter (6) performs inverse Bahr conversion on the subsequences CD, CI, ..., CM-1 of the Baar spectrum described above to obtain M step-like number sequences Fo, Fu, ...,
Calculate FM-1. Here, the degree M is equal to the number of steps of the step-like function.

First, G51 (transforming formula 161 to ψ and 1) = "2 (Co -) C1")
α9M p and tshi' (C0-01ttl)asM are obtained. However, if M=21 and sequentially calculate 191■Formula%Formula%), then FJ=ψJ=ψ'', (3=o, 1....,M-1
) *1) Numerical sequence Fo, Fl, --..., h from busy
-1 is obtained. This is the approximate digital signal FO,
Fl,..., FM-1 itself.

The expressway bar converter (6) has a function of calculating equations 09 to 01.

Next, the error calculator (7) calculates a sequence of numbers obtained from the expressway bar converter, ie, approximate digital signals Fo, Fx.

..., FM-1 and the digital signal ψ0. stored in the data buffer (3). ψ1. ..., compare ψN-1 and calculate the error car using the 41 f5+ formula. This error calculation is repeated while sequentially increasing the order M of the bar spectrum specified to the spectrum selector (5) from 1.

It is stopped when the error translation falls below the predetermined optimal error e for the first time.

The order of the Baar spectrum at this time is the optimal order M, and the sequence Fo, F which is the output of the expressway Baar converter (6)
1. -..., -t is a step-like function consisting of i steps, and the sequence FO, Fl,..., FT-t is the digital signal ψ0. ψ1. ..., approximate digital signals Fo, Fl, ..., ψ-1 that optimally approximate ψN-1
becomes.

Therefore, approximate digital signals Fo, Ft,..., n-1
The current value F'lt is outputted from the expressway bar converter (6) as a smoothed value hill.

Note that the approximate digital signals FO, Fl, . . .
The error corresponding to 1 is taken as the variance σ2 and the error calculator (7)
It is output from

Namely.

Yx = F Miya-1 (sha) σ2 = ρn (c)
It is.

Next, the operation of the prediction calculator (8) will be explained.

Digital signal ψ0 stored in data buffer (3)
If there is no pulse-like signal (1) indicating an abnormality in ゜ψl, ..., ψN-1, the above smoothed value YK represents the low frequency noise obtained by removing the high frequency noise from the measurement signal X. It is known that changes in low frequency noise can be expressed by the autoregressive model of equation (2).

Yic=AxK-1+BXx-z+・-・+CXk-x
+eK (241 (to) In formula, eK is unpredictable white noise, A, B.

..., C are constants called model parameters.

Model parameters are often unknown, and estimation calculations are performed from past values of low-frequency noise. As the simplest example (
By setting the formula as K=2, eK==Q and using the four past values of the smoothed value YK-t, ..., YK-4, the formula (251 (a) is obtained.

YK- inside AYK-2-1-BYK-3 (to) YK-2
Medium AYK-s +BYK-4-□□□) (25) r:I
Solving the S1 formula results in the following. here.

D=YK-2・YK-4-YK-3@E=YK-t −YK-4-YK-2・YK-s
@F = YK-2-YK-1-YK-3
It is CIG+. Using the model parameters A and B obtained from equation (5), the current value YK of the smoothed value can be predicted.

If this is the predicted value fK, it can be calculated using formula 739.

rpc = A YK-1-4-B YK-2 (3υ prediction calculator (8) stores the past value of the smoothed value YK (
It has the function of calculating the ~Gυ formula.

Next, the discriminator (9) uses the smoothed value hill that is the output of the expressway bar converter (6), the predicted value fK that is the output of the prediction calculator (8), and the variance σ2 that is the output of the error calculator (7). A pulse-like signal (1) indicating an abnormality included in the measurement signal X using
Detect.

Digital signal ψO stored in data buffer (3)
If there is no pulse-like signal (1) indicating an abnormality in 9ψl, ..., ψN-1 and there is only noise, the difference between the predicted value fx and the smoothed value hill is small, so 1The square of the difference between the two is If it is assumed to be direct, σ falls within the range of a times the variance σ2. However, a is a constant determined from experimental results.

Therefore, the formula (to) (to) is obtained.

aK= (rK-Yx)” Cl
2dK<aσ2 (to) On the other hand, the digital signal ψ stored in the data buffer (3)
0. ψ1. ..., if there is a pulse-like signal (1) indicating an abnormality in ψN-1, the above-mentioned autoregressive model cannot predict the pulse-like signal (1), so the above predicted value fK is a smoothed value Y
There is a big difference between him and K. Therefore aK≧aσ2
It becomes C141.

The discriminator (9) sets the detection signal dK to 0 while the formula [Yes] is satisfied, sets the detection signal dx to 1 when the formula is satisfied, and outputs a pulse-like signal (1) indicating an abnormality in the digital signal version.
) is detected and notified when it occurs.

FIG. 5 shows the smoothed value function, noise predicted value fK, variance σ2. Predicted value fK and smoothed value Yx
This figure shows the approximate shape of the square of the O difference αK and the detection signal.

As described above, according to the abnormality detection device of the present invention, the high frequency noise of the measurement signal of the vibration accelerometer is removed by an adaptive filter configured by combining the Hale transform and the inverse Var transform, and a pulse-like signal indicating an abnormality is generated. In addition, low-frequency noise is predicted using an autoregressive model, and when the prediction is incorrect, it is detected that an abnormal signal that is not noise has occurred in the measurement signal, so abnormalities can be detected without delay.

It is possible to perform more reliable diagnosis of equipment such as power generation equipment.

Although an example of detecting an abnormality in a power generation device using a vibration accelerometer is shown as an example, the present invention is not limited to a zero power generation device, but can be applied to abnormality detection in any equipment, device, or natural phenomenon. .

Furthermore, it is also possible to apply the present invention to abnormality detection of time-varying signals such as speed signals, acoustic signals, and magnetic signals, regardless of acceleration signals.

[Brief explanation of drawings]

Fig. 1 is a diagram showing a measurement signal from a vibration accelerometer attached to a power generation device, Fig. 2 is a configuration diagram showing an embodiment of an abnormality detection device according to the present invention, and Fig. 3 is an approximation of the measurement signal by a step function. FIG. 4 is a diagram showing a part of the Barr function, and FIG. 5 is a diagram explaining the shape of the smoothed value for the measurement signal, the predicted value 9 variance, the detected signal, etc. In the figure, (1) is a pulse-like signal indicating an abnormality, (2) is a ~Φ converter, (3) is a data buffer, (4) is a high-speed bar converter, (5) is a spectrum selector, and (6) is a Expressway bar converter, (7) is error calculator, (8) is prediction calculator,
(9) is a discriminator. It should be noted that the same or equivalent parts in FIG. 0 are denoted by the same reference numerals. Agent Shin Kuzuno - Figure 1 Figure 3 Figure 4 Figure 5

Claims (1)

[Claims] Inputs measurement signals obtained from various sensors. and an A/D converter that converts this into a digital signal. A data buffer that stores the values of the digital signal from the present to a certain past while updating them in synchronization with the conversion cycle of the ~Φ converter, a high-speed 7%, -7 converter, a spectrum selector, and an expressway. It consists of a Nord converter and an error calculator, and calculates a plurality of Haar spectra by Nordling the digital signal stored in the data buffer, and Reverse the following subsequence.
- Calculate an approximate digital signal that approximates the stored digital signal by converting the signal, and compare this approximate digital signal with the stored digital signal to determine if the approximation error is smaller than the predetermined optimal error, and Select the subsequence of the Nord spectrum that is closest to the current value, output the approximation error at this time as a variance, and use the current value obtained from the approximate digital signal at this time as a smoothed value, that is, the filter output. An adaptive filter that performs approximation according to the characteristics of the stored digital signal each time the stored value of the data buffer is updated and removes high frequency noise; and a prediction calculator that stores past values and calculates a predicted value of the smoothed value using an autoregressive model, and compares the predicted value, the smoothed value, and the variance to detect a sudden change in the measured signal. An abnormality detection device characterized in that it detects an abnormality or a peculiar signal that appears.