JPH03235027A - Abnormality detecting apparatus - Google Patents

Abstract

PURPOSE:To detect the change of acoustic signals separately from the stationary change accompanying the operation of plant equipments thereby to improve the detecting accuracy by identifying the catching acoustic signals with an auto-regressive model in the form of a plurality of time series data of different lengths of the storing time. CONSTITUTION:A sound processor 9 changes sound signals obtained through a microphone 3 into digital signals which are taken into an input part 10 and processed in an operating part 11. As a result, whether an equipment which is a target to be monitored is normal or abnormal is judged. The judging result is output from an output part 12 and displayed at 13. According to a self- regression analysis function, the signals are identified with an auto-regressive model 24 to estimate a predicting value 25 at a next sampling time point T+DELTAT and a variance value 26 of white noise components attributing the predicting value 25. Further, the residual between the predicting value 25 and an actually- measured value 27 after another sampling is calculated and then compared with the variance value 26. A probability when the actually-measured value 27 is regarded as outside the change at the normal time is calculated as the degree of reliability for judgement of abnormality.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention detects abnormalities in equipment and equipment in power generation plants and other various plants based on changes in acoustic signals over time.
This invention relates to an anomaly detection device that has an automatic detection function.

(Prior technology) In recent years, various plants such as power plants have been using acoustic systems to detect abnormalities that can be detected audibly, such as leakage of liquid or gas from plant equipment such as piping, malfunction of rotating machines, or normalization of piping. Various automatic detection methods using signal processing techniques have been attempted.

One of these methods will be explained using an example of steam leak detection from piping shown in FIGS. 10 and 11.

In FIG. 10, when a steam leak 2 occurs from a piping flange 1, an acoustic signal captured by a microphone 3 installed in the vicinity is processed by an acoustic processing device 4 to detect the presence or absence of a steam leak. 5 shows a control valve inserted in the piping system.

FIG. 11 shows the frequency characteristics of the spectral level in the acoustic signal captured by the microphone 3. When a steam leak occurs, the spectral level increases 7 in a certain frequency band compared to the normal spectrum 6. Therefore, by analyzing this change in the spectral level using the acoustic processing device 4, it is possible to determine whether there is an abnormality in the plant equipment.

(Problem to be solved by the invention) However, in normal large-scale plants, a huge number of devices are operated under organic continuous control, so the acoustic signals generated from a large number of devices are Spectral levels are constantly changing over time.

For example, as shown in FIG. 10, if the flow rate control valve 5 whose opening degree changes as the plant operates is close to the piping flange 1 to be monitored, the The spectral level of the acoustic signal captured by the microphone 3 changes periodically in response to changes in the opening degree of the valve 5.

Therefore, even if the pipe flange 1 to be monitored is normal, different spectra will be obtained depending on the measurement time as shown in FIG. 12.

On the other hand, since the input signal of the microphone 3 has a fluctuation range 8 in the spectral level due to the operation of the plant equipment, as shown in FIG. 7 falls within the fluctuation range 8 due to plant operation, and there is a risk that the detection accuracy will decrease.

As a measure to improve this detection accuracy, we should investigate the causal relationship between the level and period of fluctuations caused by plant operation and the operation of plant equipment that causes the fluctuations, and consider a method to correct spectral level fluctuations caused by plant operation. It will be done.

However, the causes of spectral level fluctuations are wide-ranging, ranging from relatively short-period factors such as changes in valve opening to long-period factors such as climate changes, and the number of component devices is enormous, making it difficult to implement the above improvement measures. conversion is impossible or extremely difficult.

The present invention has been made in order to solve these problems, and the present invention has been made in such a way that the change in the acoustic signal when the equipment to be monitored is abnormal is mixed with the periodic fluctuation component that occurs with the operation of the plant equipment group. We aim to provide an abnormality detection device that can detect abnormalities by processing acoustic signals even in equipment located in locations where acoustic signals fluctuate significantly due to plant operation. This is the purpose.

[Structure of the Invention] (Means for Solving the Problem) In order to achieve the above object, the abnormality detection device of the present invention detects an acoustic signal emitted from a device to be monitored to detect an abnormality in the device. In the detection device, there is an input section that has a sampling function that repeatedly takes in the acoustic signal obtained through the microphone as a digital quantity at regular intervals; Series data generation function,
The time-series data of each time length is identified using an autoregressive model, and the change process of the acoustic signal during normal times is calculated as a predicted value, and the actual measured value is determined from the residual difference between this and the actual measured value obtained by the sampling. An autoregression analysis function that calculates the probability that the data is outside the change process as the reliability of abnormality judgment, and a judgment function that compares the reliability calculated for each time series data with a threshold value to determine the presence or absence of an abnormality. an arithmetic unit having; an output unit that outputs the determination result to a reporting device;
It is characterized by comprising a sound processing device consisting of the following.

(Function) According to the abnormality detection device having such a configuration, the change in the acoustic signal at the time of equipment abnormality is prevented from being buried in periodic fluctuation components that occur with the operation of the plant equipment group. This makes it possible to detect abnormalities even in monitored equipment located in areas where the acoustic signal fluctuates significantly due to the operation of the system.

Next, the operation of the apparatus of the present invention will be explained in more detail with reference to FIGS. 1 to 4. In addition, in these figures, the first
The same parts as in the figures after figure 0 are given the same reference numerals.

FIG. 1 is a schematic diagram showing the configuration of the device of the present invention, in which an acoustic processing device 9 includes an input section 10 that digitizes and captures the acoustic signal obtained through the microphone 3, and an input section 10 that processes the digitized acoustic signal. It is comprised of an arithmetic unit 11 that determines whether the equipment to be monitored is normal or abnormal, and an output unit 12 that outputs the determination result.

Reference numeral 13 indicates a reporting device such as a CRT that displays the determination result.

FIG. 2 shows the functions of the sound processing device 9, and the sampling function 14 is a function that the input section 10 of the sound processing device 9 has. Further, the time series data generation function 15, the autoregressive analysis function 16, and the determination function 17 are functions belonging to the calculation unit 11, and the output function 18 is a function that only the output unit 12 has.

In the above, the sound processing device 9
The obtained acoustic signal is first processed by sampling function 1 in Figure 2.
4, it is repeatedly taken in as a digital quantity at fixed time intervals. In addition, the acoustic signals digitized by the time series data generation function 15 are converted into multiple time series data with different accumulation time lengths T, as shown in 19 to 21 in FIG. The data is saved sequentially at each time interval.

The autoregressive analysis function 16 first analyzes the change process of each time series data shown in 19 to 21 in FIG. 3 based on the acoustic signal level 23 actually measured at time T.
4, the predicted value 25 at T and this predicted value 25 are identified in the autoregressive model as shown in 4.
The variance value 26 of the white noise component incidental to is estimated.

Furthermore, after calculating the residual difference between the predicted value 25 and the actual value 27 obtained by performing new sampling, we compare it with the variance value 26 to determine the probability that the actual value 27 is outside the normal change process. Calculated as the reliability of the judgment.

The determination function 17 in FIG. 2 determines the presence or absence of an abnormality by comparing the reliability of the abnormality test calculated by the autoregression analysis function 16 with a threshold value.

Further, the output function 18 outputs the determination result to the notification device 13.

(Example) Next, an embodiment of the device of the present invention will be described.

FIG. 5 shows, as an example of the present invention, a configuration example of an abnormality detection device installed for the purpose of monitoring a fluid control valve 5 in a power generation plant. The sound processing device 9 held in
and a notification device 13.

The acoustic signal captured by the microphone 3 fluctuates in a complicated manner with various periods due to the sources 30 of periodically fluctuating sounds scattered around and the squeezing sound generated from the control valve 5 itself, which is the object of monitoring.

Therefore, in order to accurately detect abnormalities with respect to control valves, it is necessary to eliminate the influence of these sounds that periodically fluctuate with the operation of the plant.

In this case, the period of fluctuation is expected to range from a long one-year period due to increases and decreases in output due to climate changes to fluctuations with a period of several seconds related to adjustment control of steam pressure and rotating machine rotation speed. be done.

In the signal processing function shown in FIG. 2, the sampling function 14 digitizes and captures the acoustic signal obtained by the microphone 3 at a sampling period sufficient to reproduce the waveform even with the shortest periodic variation.

The time-series data generation function 15 holds the acoustic signals sequentially obtained as digital quantities for a certain length of time as short-time time-series data 31, as shown in FIG. Further, as the short-time time series data 31 is updated, the short time time series data is sequentially averaged to calculate and accumulate intermediate time series data 32 and long time time series data 33.

The time length of each of the time series data 31 to 33 is set so as to be able to completely capture various fluctuation cycles of the change process accompanying the operation of the plant. For example, the time length of short time series data 31 that corresponds to relatively short period fluctuations related to adjustment control of pressure etc. is set to about 10 seconds, and the time length of long time series data 33 that corresponds to relatively short cycle fluctuations related to adjustment control of pressure etc. is set to about 10 seconds. It is considered effective to set the time length to about two years.

The autoregressive analysis function 16 uses these time series data 31 to
33, the autoregressive model shown in FIG. 4 is identified.

Next, details of the procedure of identification to the autoregressive model and comparison with actual measured values will be explained.

First, time series data (Xi) of data number N (where i
−1 to N), and predicted linkage 6+ at the next measurement point N+1
1 is identified as an autoregressive model using the following equation.

Father N+ I ”” a l X N 10a 2 X N
-1+ ”'”'10am X N-s +l + e
N+4 ''''■Here, (ail (however,
i=1~m) is an autoregressive coefficient, eNm is predicted linkage N+4
This is the white noise component attached to the . Note that the autoregressive coefficient (
al), the white noise component eN+4, and the order m of the autoregressive model are determined by applying the minimum prediction error criterion (Final Prediction) proposed by Akaike to the Neil Walker equation composed of the autocorrelation function of time series data (Xi
Error, F P E ).

Next, regarding the predicted linkage Nll obtained using the formula ■,
The residual difference S N+ + with the actual measurement value X N+ + obtained by newly sampling is calculated using the following equation.

SN++=XN+1-Father N+1...
・・・・■Next, the probability that the measured value xN++ can be considered to be outside the change process in past time series data (Xi l is calculated using the following formula as the reliability P in the standard normal distribution.Judgment function 17 is the formula 0. Normality/abnormality is determined by comparing the reliability P of abnormality determination corresponding to each time series data calculated in with a threshold value.

The following table shows an abnormality determination threshold table held in the calculation unit 11.

The result determined by the determination function 17 is output to the notification device 13 via the output function 18.

In the device of the present invention configured as described above, the waveform of the acoustic signal captured by the microphone 3 can be reproduced even with the shortest cycle fluctuation by the sampling function 14 of the input section 10 inside the acoustic processing device 9. The data is digitized and captured at a sampling period set to a certain level.

Next, the time series data generation function 15 of the arithmetic unit 11 is configured so that relatively short period fluctuations are captured in short time series data 31 with a short accumulated time length, and long period fluctuations are captured in long time series data 31. 33, the acoustic signals are accumulated into three time-series data sets each having a different time length.

The autoregressive analysis function 16 identifies each time-series data to an autoregressive model using equations 1 to 0, and then calculates the reliability P of abnormality determination.

The determination function 17 compares the reliability P of abnormality determination with the abnormality determination threshold table shown in Table 1, and the reliability P of either time series data is equal to the threshold A or B.

When it reaches C, it is determined that there is an abnormality. This determination result is led to the output function 18 of the output unit 12, and outputs the distinction between normal and abnormal and the reliability of the determination to the reporting device 13.

In the above-described embodiment of the present invention, each time series is created in such a way that it captures everything from short-cycle fluctuations on a second-by-second basis to long-cycle fluctuations on a yearly basis that occur as a result of the operation of a group of plant equipment. By comparing and setting the time length of data and identifying it with an autoregressive model, it is possible to detect relatively short-period fluctuations related to adjustment control of pressure, etc., as well as long-period fluctuations such as those associated with climate changes. This makes it possible to clearly separate the change from the change in equipment abnormality and make a determination, resulting in an improvement in detection accuracy.

FIG. 7 shows another embodiment of the device of the present invention, in which the abnormality determination threshold is changed for an operation that causes a change in the periodic change process of an acoustic signal among normal plant operation operations. 3 is a functional explanatory diagram of the sound processing device 4 having a threshold value correction function 35. FIG.

When non-periodic plant operation operations such as changes in load settings are performed in a power generation plant, the previous periodic change process will transition to a new change process. This will be explained with reference to FIG.

In contrast to the change process of the sound signal level before the execution of the driving operation shown in 36 in Figure 8, after the time 37 of the driving operation, 38
Assuming that the change process shown in has occurred, real/1!
The residual S between the 1 value X and the predicted value RN+1 is 40 in Figure 8.
As shown in , immediately after the driving operation execution time point 37,
This is a sharp increase from the previous residual of 39. Then, the predicted linkage N
Since +1 gradually and accurately estimates the change process 38 after the driving operation is performed, the residual S gradually decreases as shown at 40.

On the other hand, as shown at 42 in FIG. 8, the magnitude of the white noise component e attached to the predicted linkage N++ is as follows:
The maximum value is reached when a considerable amount of 8 data is connected.

As a result, the reliability P of abnormality determination calculated by the autoregressive analysis function 16 in FIG. , as shown at 44 in FIG. 8, is highest immediately after the driving operation execution point 37, and then decreases as the residual S decreases and the white noise component e increases.

By performing such an aperiodic driving operation, there is a risk that the reliability P of the abnormality determination exceeds the threshold value 45 during a certain period of time immediately after the driving operation execution time point 37, resulting in an erroneous determination.

Therefore, in the embodiment of the present invention shown in FIG. 7, the threshold correction function 35 obtains information indicating the presence or absence of non-periodic driving operations from a plant control computer or the like.

In this case, as shown in Table 2, the operation contents are defined in advance in the sound processing device as a plant data setting table, and the information is acquired when the operation set in this table is performed.

Table 2 Plant Data Setting Table Next, the threshold correction function 35 sets the abnormality determination threshold value in advance as shown in 47 in FIG. For the abnormality judgment threshold 45 set in the table (Table 1),
A value proportional to the residual S is added to correct it. In this case, correction is performed for the data interval until the residual S after the plant operation converges to a certain value or less.

The determination function 17 in FIG. 7 determines the presence or absence of an abnormality by comparing the corrected abnormality determination threshold 47 with the reliability P of abnormality determination.

According to this embodiment, information regarding the presence or absence of operation of plant operating equipment that affects the periodic change process of acoustic signals is obtained and the abnormality determination threshold is corrected, thereby preventing non-periodic changes. This has the effect of preventing erroneous judgments caused by driving operations.

[Effects of the Invention] As described above, according to the present invention, by identifying the captured acoustic signals in the form of multiple time series data with different accumulation time lengths to an autoregressive model and determining the presence or absence of an abnormality, Detects changes in acoustic signals caused by equipment abnormalities by separating them from steady fluctuation components, ranging from short-period components on the order of seconds to long-period components on the order of years, which occur with the operation of plant equipment. This improves detection accuracy, and furthermore, as the power spectrum fluctuation pattern is automatically learned in the process of continuing identification with the autoregressive model, differences in detection accuracy due to the ability of individual coordinators are less likely to occur. Effectiveness is obtained.

[Brief explanation of drawings]

FIG. 1 is a schematic diagram showing the configuration of the device of the present invention, and FIG. 2 is a flowchart showing the functions of the sound processing device of the present invention.
Fig. 3 is a graph illustrating a plurality of time series data with different time lengths, Fig. 4 is a graph illustrating an identification method for an autoregressive model for time series data, and Fig. 5 is a configuration of an embodiment of the device of the present invention. FIG. 6 is a graph illustrating three sets of time series data with different time lengths, FIG. 7 is a flowchart showing the functions of the sound processing device in another embodiment of the present invention, and FIG. A graph explaining the transition of the change process due to plant operation, FIG. 9 is a graph showing the abnormality judgment threshold after correction, FIG. 10 is a schematic diagram explaining the conventional detection method for steam leaks from piping, and FIG. Figure 11 is a graph that explains the increment in the spectrum due to steam leakage from piping, Figure 12 is a graph that explains the difference in spectrum depending on the measurement time, and Figure 13 is a graph that explains the fluctuation of the spectrum due to the operation of plant equipment. be. 3... Microphone 4.9... Sound processing device 10... Input section 11... Arithmetic section 12... ...Output section 13...Notification device

Claims (1)

[Claims] In an abnormality detection device that detects an acoustic signal emitted from a device to be monitored and detects an abnormality in the device, sampling is performed to repeatedly capture the acoustic signal obtained through a microphone as a digital quantity at regular intervals. An input section with functions: a time-series data generation function that stores sampled acoustic signals as multiple time-series data with different time lengths, and an autoregressive model that identifies the time-series data of each time length to determine normal times. After calculating the change process of the acoustic signal as a predicted value, the probability that the actual value can be considered to be outside the normal change process is calculated from the residual difference between this and the actual value obtained by the sampling as the reliability of abnormality determination. an arithmetic unit that has an autoregressive analysis function that performs the calculation, and a determination function that compares the reliability calculated for each time series data with a threshold value to determine the presence or absence of an abnormality; an output unit that outputs the determination result to the reporting device. An anomaly detection device comprising: an acoustic processing device comprising; and;