JP4845695B2 - Anomaly detection apparatus and method - Google Patents

Description

  The present invention relates to an abnormality detection apparatus and method for detecting an abnormality of a monitoring target by detecting sound generated in the monitoring target.

  In facilities where combustible gas is handled or used, it is necessary to detect leakage of combustible gas as an abnormality and detect it with high accuracy. Such facilities include hydrogen handling facilities such as hydrogen stations and hydrogen supply plants, natural gas handling facilities, petroleum gas handling facilities, liquid fuel handling facilities, and facilities that use these gases.

  Conventionally, when detecting an abnormality in a monitoring target, a detection unit is provided for the monitoring target, and a predetermined threshold is set based on an output from the detection unit when the monitoring target is in a normal state in advance. Abnormality was detected by comparing the measured value by and the set threshold value.

Further, in Patent Document 1, in order to accurately determine an abnormality even when a measured value changes gradually or suddenly, one type of measured value is checked from various angles, and the result is mainly used. There has been proposed an abnormal value detection apparatus that performs comprehensive determination using component analysis.
JP-A-10-260722

  However, in the case where gas leakage is detected as abnormal by detecting sound in the monitoring target facility, it may be erroneously determined to be abnormal although there is no gas leakage and is normal. That is, in the equipment to be monitored, in addition to the gas leak sound that should be detected as an abnormality, there are work sounds such as air cleaning similar to the gas leak sound. Such a working sound causes a false alarm that is erroneously determined to be abnormal. Therefore, in order to realize high-accuracy abnormality detection, it is necessary to journal gas leakage sound that should be detected as abnormality and other work sounds.

  An object of the present invention is to detect a gas leakage sound that should be detected as an abnormality by using a work sound similar to the gas leakage sound and detecting it with high accuracy, and detecting an abnormality with high accuracy. And providing a method.

The above object is achieved by a sound collector to output a sound signal by detecting an acoustic in the monitored hydrogen gas is hydrogen station to be handled, the calculated frequency spectrum of the acoustic signal, and advance the current normal sound clef Using a neural network model in which abnormal sound is detected and the normal sound data output is 0 and the abnormal sound data output is 1 , the normal sound is detected based on the frequency spectrum. A frequency pattern analysis unit that outputs a judgment value quantified to a numerical value of 0 to 1 according to normal to abnormal , with a judgment value of 0 when the hydrogen gas leak sound is detected and a judgment value of 1 when the hydrogen gas leakage sound is detected For the time-series data of the determination value output for 5 to 30 seconds , an average value and a standard deviation of the determination value are calculated, and the calculated average value and the first threshold value are calculated. And the calculated standard deviation is compared with a second threshold value, and the monitoring is performed when the average value is equal to or greater than the first threshold value and the standard deviation is equal to or smaller than the second threshold value. It is achieved by an abnormality detection device comprising: an abnormality detection unit that determines that the state of an object is abnormal and detects leakage of the hydrogen gas as the abnormality.

In addition, the above-mentioned purpose is to detect sound in a monitoring target that is a hydrogen station where hydrogen gas is handled, generate an acoustic signal, calculate a frequency spectrum of the acoustic signal , detect normal sound and abnormal sound in advance, Using a neural network model that has been trained with an output of sound data of 0 and an output of abnormal sound data of 1 , based on the frequency spectrum, the state of the monitoring target is determined as a determination value when a normal sound is detected. The determination value when the hydrogen gas leakage sound is detected is set to 1 , and a determination value that is quantified to a numerical value of 0 to 1 according to normal to abnormal is generated, and the determination value generated for 5 to 30 seconds For the time-series data, an average value and a standard deviation of the determination values are calculated, the calculated average value is compared with a first threshold value, and the calculated standard deviation and the second standard deviation are compared. A comparison with a threshold value, and when the average value is not less than the first threshold value and the standard deviation is not more than the second threshold value, it is determined that the state of the monitoring target is abnormal, and the hydrogen gas This is achieved by an abnormality detection method characterized by detecting a leakage of the above as the abnormality.

  According to the present invention, sound is detected in a monitoring target to generate a sound signal, a frequency spectrum of the sound signal is calculated, a model of a neural network is used, and the state in the monitoring target is determined based on the frequency spectrum. Gas to be detected as an abnormality because a determination value converted into a numerical value corresponding to normal to abnormality is generated and the abnormality of the monitoring target is detected based on time-series data of the determination value generated during a predetermined period It is possible to detect and detect an abnormal leak with high accuracy by using a work noise similar to that of the gas and detecting it with high accuracy.

[One Embodiment]
An abnormality detection apparatus and method according to an embodiment of the present invention will be described with reference to FIGS. FIG. 1 is a schematic diagram illustrating a configuration of an abnormality detection device according to the present embodiment. FIG. 2 is a typical example of time-series data of determination values output by a neural network determination unit in the abnormality detection device of the abnormality detection device according to the present embodiment. FIG. 3 is a graph showing a result of trend processing performed on the time-series data of determination values by the trend processing unit in the abnormality detection device according to the present embodiment. FIG. 4 is a graph of the abnormality detection unit in the abnormality detection device according to the present embodiment. FIG. 5 is a graph showing time series data of determination values obtained for the hydrogen gas leakage sound in the abnormality detection method according to the present embodiment. FIG. 6 is a diagram showing the air cleaning work sound in the abnormality detection method according to the present embodiment. FIG. 7 is a graph showing the time-series data of the obtained determination values. FIG. 7 shows the train in the abnormality detection method according to this embodiment. Is a graph showing the results of the processing.

  The abnormality detection device according to the present embodiment detects leakage of hydrogen gas as an abnormality in a monitoring target facility that handles hydrogen, such as a hydrogen production plant or a hydrogen station.

  As shown in FIG. 1, the abnormality detection apparatus according to the present embodiment includes a sound collection unit 10, a frequency pattern analysis unit 12, and an abnormality detection unit 14.

  The sound collecting unit 10 includes a microphone unit 16 that detects sound generated in the monitoring target facility and outputs an acoustic signal, an analog amplifier unit 18 that amplifies the acoustic signal output from the microphone unit 16, and an analog amplifier unit 18. And an A / D converter 20 that converts the amplified acoustic signal from an analog signal to a digital signal.

  The frequency pattern analysis unit 12 includes a Fourier transform unit 22 that calculates a frequency spectrum of the acoustic signal converted into a digital signal by the A / D conversion unit 20, and a neural network based on the frequency spectrum calculated by the Fourier transform unit 22. And a neural network determination unit 24 that outputs a determination value corresponding to the sound according to the determination model according to.

  The abnormality detection unit 14 includes a trend processing unit 26 that obtains a trend of the determination value output from the neural network determination unit 24, and an identification unit 28 that identifies hydrogen gas leakage sound and work sound based on the result of the trend processing unit 26. have.

  First, each component of the sound collection unit 10 will be described.

  The microphone unit 16 is installed in the monitoring target facility, detects sound generated in the monitoring target facility, and outputs an acoustic signal. The microphone unit 16 is non-directional having sensitivity, for example, at 20 Hz to 20 kHz. The microphone unit 16 has an intrinsically safe explosion-proof structure.

  The analog amplifier unit 18 amplifies the acoustic signal output from the microphone unit 16 with a predetermined gain. The analog amplifier unit 18 amplifies an acoustic signal of 20 Hz to 20 kHz, for example, with which the microphone unit 16 has sensitivity.

  The A / D conversion unit 20 converts the acoustic signal amplified by the analog amplifier unit 18 from an analog signal to a digital signal at a predetermined sampling frequency. The resolution of A / D conversion by the A / D converter 20 is, for example, 8 to 16 bits, and the sampling frequency is selected from, for example, 44.1 kHz, 48.0 kHz, and 51.2 kHz.

  Next, each component of the frequency pattern analysis unit 12 will be described.

  The Fourier transform unit 22 performs a fast Fourier transform on the acoustic signal converted into a digital signal by the A / D conversion unit 20 and calculates a frequency spectrum of the acoustic signal. The number of data points of the fast Fourier transform by the Fourier transform unit 22 is selected from 512, 1024, 2048, 4096, 8192, 16384, 32768, and 65536, for example. The window function is selected from, for example, hamming, hanning, and no window function.

  The neural network determination unit 24 outputs a determination value corresponding to the sound detected in the monitoring target facility based on the frequency spectrum calculated by the Fourier transform unit 22 using the neural network model.

  In the learning of the neural network, the normal sound data is output as 0 and the abnormal sound data is output as 1. As the neural network, a hierarchical neural network is used, and learning is performed by back propagation. The response function in the neural network is selected from, for example, a sigmoid function and a radial base function.

  Using the learned neural network model, the neural network determination unit 24 outputs a determination value of 0 to 1 according to the sound detected in the monitoring target facility. Specifically, when a hydrogen gas leakage sound is detected, the determination value 1 is output. When a normal sound is detected, a determination value 0 is output. Further, a determination value closer to 1 is output as the detected sound is more similar to the hydrogen gas leakage sound, and a determination value closer to 0 is output as it is similar to the normal sound. While the monitoring target equipment is being monitored, the neural network determination unit 24 outputs such a determination value at a predetermined cycle.

  It is considered that when a hydrogen gas leak sound is detected, the determination value 1 is output almost continuously, and when the hydrogen gas leak sound changes due to wind or the like, a determination value close to 1 is output. Note that the hydrogen gas leakage sound does not vary greatly as long as there is no external factor such as wind. Therefore, it is considered that the frequency at which the determination value becomes smaller than 1 is low.

  On the other hand, the air cleaning operation sound generated during the operation of cleaning the vehicle body with air is similar to the hydrogen gas leakage sound. For this reason, when an air cleaning work sound is detected, it is considered that a determination value close to 1 or 1 is output. However, unlike the hydrogen gas leak sound, the air cleaning work sound is generated along with the work by the worker, and is generated irregularly in time. Accordingly, the determination value 1 or a value close to 1 based on the detection of the air cleaning work sound is not output continuously, and if the normal sound is detected, the determination value close to 0 or close to 0 is output with high frequency. Conceivable.

  FIG. 2 is a table showing a typical example of time-series data of determination values output by the neural network determination unit 24. In the figure, No. 1-No. 3 shows time-series data of judgment values when hydrogen gas leakage sound is detected. 3-No. 6 shows time-series data of determination values when air cleaning work sound is detected. No. 3-No. The air cleaning work sound in 6 is a work sound generated during the work of blowing air to the vehicle body with an air gun for cleaning.

  If a hydrogen gas leak sound is detected, no. 1-No. In any of the cases 3, the determination value 1 is output almost continuously from the neural network determination unit 24, and the determination value 0.7 or 0.8 close to 1 is occasionally output once or twice. For this reason, the average value of the judgment values during this period is as large as 0.95 to 0.98, and the standard deviation is as extremely small as 0.06 to 0.11.

  On the other hand, when an air cleaning work sound is detected, No. 3-No. In any of the cases 6, when the air gun is far from the vehicle body and the air blowing sound does not increase, the determination value 0.3 is output, and when the air gun approaches the vehicle body and the blowing sound increases, the determination value 1 is set. It is output. Since the work by the worker is irregular in time, when the determination value 1 is output, it is 2 to 5 times, and the determination value 1 is not continuously output.

  In this way, the frequency pattern analysis unit 12 generates time-series data of determination values obtained by quantifying the state of the monitoring target facility into a numerical value corresponding to normal to abnormal.

  Next, each component of the abnormality detection unit 14 will be described.

  The trend processing unit 26 performs trend processing on the time-series data of the determination values generated in the frequency pattern analysis unit 12. Specifically, the trend processing unit 26 calculates an average value and a standard deviation of the determination values for the time series data of the determination values.

  FIG. 3 shows time-series data Nos. Of determination values shown in the table of FIG. 1-No. 6 is a graph in which an average value and a standard deviation of judgment values are calculated for each of 6 and plotted with an average value on the horizontal axis and a standard deviation on the vertical axis. In the table of FIG. 1-No. 6 shows the calculated average value and standard deviation.

  As is apparent from the graph shown in FIG. 3, the hydrogen gas leakage sound and the air cleaning work sound are plotted in different regions partitioned by an average threshold value a and a standard deviation threshold value b. In this way, the hydrogen gas leakage sound and the air cleaning work sound can be distinguished from each other by the average threshold value a and the standard deviation threshold value b.

  Based on the result of the trend processing by the trend processing unit 26, the identification unit 28 can change the state of the monitored equipment to “normal” which is a normal state, “gas leakage” in which hydrogen gas is leaking, and hydrogen gas leakage It is determined that it is either “Caution” that has a characteristic or “Air cleaning operation” in which an air cleaning operation is performed. Hereinafter, the processing of the trend processing unit 26 and the identification unit 28 in the abnormality detection unit 14 will be described with reference to FIG.

  First, for the time-series data (step S10) of the determination value x generated by the frequency pattern analysis unit 12 in a predetermined period, the trend processing unit 28 sets the average value Ave (x) of the determination value x and the determination value x. A standard deviation σ is calculated (steps S12 and S14). The period of the time-series data of the determination value x for which the average value Ave (x) and the standard deviation σ should be calculated is, for example, 5 to 30 seconds, and the number of data is 5 to 30 when every other second.

  Next, the identification unit 28 determines whether or not the average value Ave (x) is greater than or equal to the threshold value a (step S16).

  If it is determined in step S16 that the average value Ave (x) is equal to or greater than the threshold value a, the identification unit 28 subsequently determines whether or not the standard deviation σ is equal to or less than the threshold value b (step S18).

  When it determines with standard deviation (sigma) being below threshold value b in step S18, the identification part 28 determines the state of the monitoring object installation as "gas leak" (step S20).

  If it is determined in step S18 that the standard deviation σ is not less than or equal to the threshold value b, that is, the standard deviation σ is greater than the threshold value b, the identification unit 28 determines that the state of the monitoring target facility is “caution” (step S22). .

  On the other hand, when it is determined in step S16 that the average value Ave (x) is not equal to or greater than the threshold value a and the average value Ave (x) is smaller than the threshold value a, the identification unit 26 subsequently determines that the standard deviation σ is smaller than the threshold value b. It is determined whether it is small (step S24).

  When it is determined in step S24 that the standard deviation σ is smaller than the threshold value b, the identification unit 28 determines that the state of the monitoring target facility is “normal” (step 26).

  When it is determined in step S24 that the standard deviation σ is not smaller than the threshold value b, that is, the standard deviation σ is greater than or equal to the threshold value b, the identification unit 28 determines the state of the monitoring target equipment as “air cleaning work”. (Step S28).

  In this way, the abnormality detection unit 14 determines that the state of the monitoring target facility is one of “gas leakage”, “caution”, “normal”, and “air cleaning work”. If it is not particularly necessary to detect that the air cleaning work is being performed in the monitoring target facility, it may be determined as “caution” instead of the determination as “air cleaning work”. (Step S28). Thus, the abnormality detection unit 14 detects leakage of hydrogen gas in the monitoring target facility as an abnormality.

  The determination result by the abnormality detection unit 14 is notified by a display device, an alarm device (not shown), or the like.

  Next, an example of an abnormality detection method using the abnormality detection apparatus according to the present embodiment will be described.

  As the microphone section 16, a 1/2 inch electrolet condenser microphone having an intrinsically safe explosion-proof structure was used together with a safety holder. The frequency range of the measurement sound was 20 Hz to 20 kHz which is an audible range.

  The measured acoustic signal data is amplified with a predetermined gain by the analog amplifier 18 and then converted from analog data to digital data by A / D conversion at an A / D converter 20 with a resolution of 16 bits and a sampling frequency of 51.2 kHz. Converted to.

  The acoustic signal converted into digital data was subjected to fast Fourier transform by the Fourier transform unit 22 with a frequency of 20 Hz to 20 kHz using Hanning as a window function and 8192 data points. Thus, the frequency spectrum was calculated for the measurement sound.

  For learning of the neural network model used in the neural network determination unit 24, the following data was used as normal sound data and abnormal sound data.

  First, as normal sound data, 40 vehicle passing sounds were used.

  As the abnormal sound data, gas leakage sound was used. Assuming piping used in the hydrogen station, gas leakage sound was generated by drilling a hole of Φ1 in a 3/8 inch piping and leaking gas from this hole. Abnormal sound data is a frequency spectrum collected as a normal sound from a frequency spectrum of a hydrogen gas leak sound of 0.9 MPa, a frequency spectrum of a helium gas leak sound of 0.9 MPa, and a frequency spectrum of a helium gas leak sound of 2.5 MPa. Was created by artificially synthesizing sound. 120 data of abnormal sound was created.

  A radial base function was used as the response function of the neural network model. The neural network model was learned with 0 as normal and 1 as abnormal.

  Using the above-described abnormality detection device, the hydrogen gas leakage sound and the air cleaning work sound were determined. In addition, about the air washing work sound, it determined about the thing in a gas station.

  FIG. 5 is a graph showing time-series data of determination values obtained for hydrogen gas leakage sound. As apparent from the graph, in the time-series data of the determination values obtained for the hydrogen gas leakage sound, the determination value close to 1 is output almost continuously.

  On the other hand, FIG. 6 is a graph showing time-series data of determination values obtained for the air cleaning work sound. The air cleaning work sound is a sound generated when a worker works. For this reason, although the air cleaning work sound is similar to the gas leakage sound, it greatly changes with time with the work. When such an air cleaning work sound is judged by a neural network model which is a non-linear response function, the change is further emphasized. That is, as is apparent from the graph shown in FIG. 6, in the time-series data of the determination values obtained for the air cleaning work sound, the determination values greatly change with time.

  FIG. 7 is a graph showing the result of trend processing performed on the time-series data of the determination values obtained for the hydrogen gas leakage sound and the time-series data of the determination values obtained for the air cleaning work sound. In the trend processing, for the time series data of the judgment value for the hydrogen gas leakage sound and the time series data of the judgment value for the air cleaning work sound, the average value and standard deviation of the judgment value in a predetermined period are obtained, respectively. The average value was plotted on the vertical axis, and the standard deviation was plotted on the vertical axis. The average value and the standard deviation were each determined from 10 judgment value data in the past 10 seconds. The period for which the average value and the standard deviation were obtained was changed along with the trend, and the average value and the standard deviation were obtained for each period.

  As can be seen from the graph shown in FIG. 7, the hydrogen gas leakage sound and the air cleaning work sound are plotted in areas that are largely different from each other, which are partitioned by the threshold value a of the average value and the threshold value b of the standard deviation. That is, the average value on the horizontal axis of the graph is x and the standard deviation on the vertical axis is y, and the hydrogen gas leakage sound is plotted in a region satisfying x ≧ a and y ≦ b. Moreover, the air cleaning work sound was plotted in a region satisfying x <a and y> b. Although not shown in the graph, normal sounds are plotted in a region satisfying x <a and y <b.

  As described above, by performing the trend processing using the average value and the standard deviation for the time series data of the judgment values based on the neural network model, the hydrogen gas leakage sound is clearly journaled as the air cleaning work sound similar to this. It was confirmed that it was detected.

  As described above, according to the present embodiment, sound is detected in a monitoring target to generate an acoustic signal, a frequency spectrum of the acoustic signal is calculated, and a state of the monitoring target is calculated based on the frequency spectrum using a neural network model. Generates a judgment value that is digitized from normal to abnormal, calculates the average value and standard deviation for the time series data of the judgment value generated during a predetermined period, and compares the calculated average value with the threshold value of the average value Since the abnormality of the monitoring target is detected by comparing the calculated standard deviation with the standard deviation threshold, the gas leakage sound that should be detected as an abnormality is compared with the work sound similar to the gas leakage sound. Journals can be detected with high accuracy.

[Modified Embodiment]
The present invention is not limited to the above embodiment, and various modifications can be made.

  For example, in the above-described embodiment, the present invention is applied when the leakage sound of hydrogen gas in a monitoring target facility that handles hydrogen such as a hydrogen production plant or a hydrogen station is detected as abnormal, but the scope of the present invention is limited to this. Is not to be done. The present invention can be widely applied to, for example, detecting a gas leakage sound in an facility that handles natural gas, petroleum gas, or the like as an abnormality.

  Moreover, although the said embodiment demonstrated the case where the leakage sound of the gas in a monitoring object facility was detected as abnormality, this invention can be applied widely when detecting abnormality of various monitoring objects based on an acoustic. .

It is a block diagram which shows the structure of the abnormality detection apparatus by one Embodiment of this invention. It is a table | surface which shows the typical example of the time series data of the determination value output by the neural network determination part in the abnormality detection apparatus by one Embodiment of this invention. It is a graph which shows the result of having performed the trend process about the time series data of the judgment value by the trend process part in the abnormality detection apparatus by one Embodiment of this invention. It is a flowchart which shows the process of the abnormality detection part in the abnormality detection apparatus by one Embodiment of this invention. It is a graph which shows the time series data of the judgment value acquired about the hydrogen gas leak sound in the abnormality detection method by one Embodiment of this invention. It is a graph which shows the time series data of the judgment value obtained about the air washing work sound in the abnormality detection method by one embodiment of the present invention. It is a graph which shows the result of the trend process in the abnormality detection method by one Embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Sound collecting part 12 ... Frequency pattern analysis part 14 ... Abnormality detection part 16 ... Microphone part 18 ... Analog amplifier part 20 ... A / D conversion part 22 ... Fourier transform part 24 ... Neural network determination part 26 ... Trend processing part 28 ... Identification part

Claims (2)

A sound collection unit that detects sound in a monitoring target that is a hydrogen station where hydrogen gas is handled and outputs an acoustic signal;
Using a neural network model in which the frequency spectrum of the acoustic signal is calculated , the normal sound and abnormal sounds are detected in advance by the sound collection unit, and the normal sound data output is 0 and the abnormal sound data output is 1. Based on the frequency spectrum, the state in the monitoring target is determined from normal to abnormal with a determination value of 0 when a normal sound is detected and a determination value of 1 when a hydrogen gas leakage sound is detected . A frequency pattern analysis unit that outputs a judgment value quantified to a numerical value of 0 to 1,
For the time-series data of the determination value output for 5 to 30 seconds , the average value and the standard deviation of the determination value are calculated, the calculated average value is compared with the first threshold value, and the calculated When the standard deviation is compared with the second threshold, and the average value is equal to or greater than the first threshold and the standard deviation is equal to or smaller than the second threshold, the state of the monitoring target is abnormal An abnormality detection unit comprising: an abnormality detection unit that determines and detects leakage of the hydrogen gas as the abnormality. Detects sound and generates an acoustic signal in the monitoring target, which is a hydrogen station where hydrogen gas is handled,
A frequency spectrum of the acoustic signal is calculated, a normal sound and an abnormal sound are detected in advance, and a model of a neural network in which normal sound data output is 0 and abnormal sound data output is 1 is used. based, wherein the condition in the monitored as a determination value when a normal sound is detected 0, the determination value when the hydrogen gas leak sound is detected 1, 0-1 numerical values corresponding to up to abnormal from normal To generate a numerical judgment value
For the time-series data of the determination value generated for 5 to 30 seconds , the average value and standard deviation of the determination value are calculated, the calculated average value is compared with the first threshold value, and the calculated When the standard deviation is compared with the second threshold, and the average value is equal to or greater than the first threshold and the standard deviation is equal to or smaller than the second threshold, the state of the monitoring target is abnormal An abnormality detection method comprising: determining and detecting leakage of the hydrogen gas as the abnormality.

Images