Abnormality predictor detection method and apparatus

Abstract

To easily detect a sign of abnormality occurrence in an object system when measurement data from the object system largely fluctuate even in a normal time or when the measurement data are sampled by a high sampling frequency.SOLUTION: A frequency spectrum is acquired by performing Fourier transform or the like on measurement data (step 101), a portion belonging to a frequency band is extracted from the frequency spectrum for each frequency band on the basis of one or more predetermined frequency bands to obtain a partial spectrum (step 102), a coordinate value of a centroid of the partial spectrum is obtained and accumulated as a data vector (step 103), and an abnormality degree (for example, a Mahalanobis distance) of a data vector corresponding to a short processing period is calculated for each processing unit period on the basis of the accumulated data vector (step 105).SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a method and apparatus for detecting a sign of an abnormality occurring in a target system. [Background technology]

[0002] Water treatment systems, such as pure water production systems that produce pure water or ultrapure water from raw water and wastewater treatment systems that treat wastewater, are configured by combining various devices, such as filtration systems, activated carbon devices, ion exchange devices, reverse osmosis membrane devices, ultraviolet oxidation devices, coagulation sedimentation devices, and electrodeionization (EDI) devices, and further include piping, heat exchangers, pumps, and the like. Patent Document 1 (Patent Document 1) discloses an example of such a water treatment system, an ultrapure water production system equipped with a booster pump for supplying the produced ultrapure water to a point of use. Since abnormalities may occur in the devices that make up such water treatment systems, it is important to detect phenomena that are predictive of the occurrence of abnormalities and to estimate the factors that may be causing the abnormalities during operation of the water treatment system. For example, pumps may exhibit predictive phenomena, such as increased vibrations of specific frequency components or the generation of specific sounds, prior to failure. Therefore, it is desirable to detect such vibrations or sounds to detect signs of pump abnormalities. Pipes may also exhibit predictive symptoms, such as increased vibration amplitude, prior to rupture or other problems.

[0003] A well-known method for detecting signs of abnormalities in various systems, including water treatment systems, is to place sensors within the system and perform multivariate analysis on the measurements obtained by the sensors. When detecting signs of abnormalities in pumps or piping, various sensors, such as vibration sensors and acoustic sensors, are placed within the system, including the pumps and piping, and ranges of values ​​that the detection values ​​of these sensors should take are set based on design values ​​based on specifications and past operating data. Whether or not the actual detection values ​​deviate from the set ranges, and if there is a deviation, the frequency of such deviations, are used to determine whether or not there are signs of abnormalities.

[0004] Patent Document 2 discloses a method for detecting signs of abnormalities in a refrigeration system using helium gas, in which measurements of multiple different parameters representing the status of at least one of the refrigerator and compressor are acquired and multivariate analysis is performed on those measurements. Patent Document 2 also discloses the use of the Mahalanobis-Taguchi (MT) system for multivariate analysis. Patent Document 3 discloses a method for acquiring time-series data on multiple process variables in a water treatment system, generating anomaly detection data from a criterion for determining whether an abnormality exists and the time-series data, diagnosing the presence or absence of an abnormality, and, if an abnormality is determined to exist, making a rule-based judgment to identify the cause of the abnormality. Patent Document 3 also discloses that the Mahalanobis distance can be used to generate anomaly detection data.

[0005] Patent Document 4 discloses collecting measurement data for multiple items from a monitored system, such as a sewage treatment system, and calculating an evaluation index through statistical processing of the measurement data and normal data collected at a predetermined time in the past. The statistical calculation of the evaluation index uses the Mahalanobis distance between the principal component scores of normal data obtained by principal component analysis and the principal component scores of the measurement data. Patent Document 5 discloses a method for calculating the cleaning cycle of a water-cooled water system of a refrigeration and air conditioning equipment by defining a reference space based on water quality data of the water-cooled water system of the refrigeration and air conditioning equipment already determined to be fouling-resistant, calculating the Mahalanobis distance from the reference space for the water quality data of the water-cooled water system of the refrigeration and air conditioning equipment being evaluated, and determining whether the refrigeration and air conditioning equipment being evaluated is fouling-resistant or fouling-prone. Patent Document 6 discloses a method for diagnosing the corrosivity of water to metals in a system having metal-water contact points, in which the Mahalanobis distance of the water to be diagnosed is calculated using data on non-corrosive water as a reference data set, with variables such as pH, conductivity, and various ion concentrations, and the corrosivity of the water to be diagnosed is evaluated based on the Mahalanobis distance value. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2024-36946 [Patent Document 2] Japanese Patent Application Laid-Open No. 2015-61993 [Patent Document 3] Japanese Patent Application Laid-Open No. 2013-61853 [Patent Document 4] Japanese Patent Application Laid-Open No. 2011-8403 [Patent Document 5] Japanese Patent Application Laid-Open No. 2006-349230 [Patent Document 6] Japanese Patent Application Laid-Open No. 2003-75325 Summary of the Invention [Problem to be solved by the invention]

[0007] As shown in Patent Documents 2-6, Mahalanobis distances are calculated based on various measurement data obtained from water treatment systems to detect signs of abnormalities in the water treatment systems. However, in water treatment systems, the water supplied to the system and subjected to treatment, i.e., raw water, can fluctuate significantly in quality. When such fluctuations occur in the raw water, the measurement data obtained from the water treatment system can fluctuate significantly even when each device in the water treatment system is operating normally. When measurement data fluctuates significantly under normal conditions, even measurement data indicating signs of abnormalities is buried under the normal measurement data, making it difficult to detect signs of abnormalities even using Mahalanobis distance. Furthermore, water treatment systems often operate continuously for long periods of time, and detecting signs of abnormalities based on the massive amount of measurement data generated by the water treatment system during such periods presents a problem: the amount of computation required is enormous. In particular, vibration sensors and acoustic sensors acquire measurements at high sampling frequencies, resulting in a particularly large amount of information, which increases the load on analytical and computational processes. It is not easy to monitor a target system in real time using vibration or acoustic sensors.

[0008] The object of the present invention is to provide an abnormality sign detection method and an abnormality sign detection device that can easily detect signs of an abnormality occurring in a target system, even when the measurement data from the target system fluctuates greatly even under normal conditions or when the measurement data is sampled at a high sampling frequency. [Means for solving the problem]

[0009] One aspect of the present invention provides an abnormality sign detection method for detecting signs of an abnormality occurring in a target system, and includes a spectrum acquisition step of converting measurement data obtained from the target system into a frequency domain representation to acquire a frequency spectrum, and a centroid calculation step of extracting, for each frequency band, a portion belonging to that frequency band from the frequency spectrum based on one or more predetermined frequency bands to obtain a partial spectrum, determining the coordinate values ​​of the center of gravity of the partial spectrum, and calculating a data vector containing at least the coordinate values ​​as elements.The spectrum acquisition step and the centroid calculation step are repeatedly performed to accumulate a series of data vectors, and the degree of abnormality of each data vector is calculated.

[0010] An abnormality sign detection device according to one aspect of the present invention is an abnormality sign detection device that detects signs of an abnormality occurring in a target system, and includes: a conversion calculation unit that converts measurement data obtained from the target system into a frequency domain representation to obtain a frequency spectrum; a centroid calculation unit that extracts, for each frequency band based on one or more predetermined frequency bands, a portion belonging to that frequency band from the frequency spectrum to obtain a partial spectrum and calculates the coordinate values ​​of the centroid of the partial spectrum; a data storage unit that stores a data vector including, as elements, the coordinate values ​​calculated by the centroid calculation unit; and an abnormality degree calculation unit that calculates the degree of abnormality of the data vector based on a series of data vectors stored in the data storage unit. [Effects of the Invention]

[0011] According to the present invention, an abnormality sign detection method and an abnormality sign detection device are provided that can easily detect signs of an abnormality occurring in a target system even when the measurement data from the target system fluctuates greatly even under normal conditions or when the measurement data is sampled at a high sampling frequency. [Brief explanation of the drawings]

[0012] [Figure 1] 1A and 1B are diagrams illustrating the principle of an abnormality sign detection method. [Figure 2]1 is a flowchart illustrating an abnormality sign detection method according to an embodiment. [Figure 3] 1 is a block diagram showing an example of the configuration of an abnormality sign detection device; [Figure 4] FIG. 1 is a flow diagram illustrating an example of the configuration of a water treatment system. [Figure 5] 10 is a graph showing an example of Mahalanobis distance calculated based on measurement data from a water treatment system. DETAILED DESCRIPTION OF THE INVENTION

[0013] Next, an embodiment of the present invention will be described with reference to the drawings. First, the principle of an abnormality sign detection method according to an embodiment of the present invention will be described.

[0014] FIG. 1 is a diagram illustrating the principle of the anomaly sign detection method of this embodiment. This anomaly sign detection method detects signs of an anomaly occurring in a target system. While the target system is not particularly limited, this anomaly sign detection method of this embodiment is preferably used when detecting signs of an anomaly occurring in a water treatment system such as a pure water production system or a wastewater treatment system. This anomaly sign detection method acquires measurement data from the target system, such as vibrations and sounds, sampled at a high sampling frequency, or measurement data that fluctuates significantly even under normal conditions. FIG. 1(a) is a graph showing an example of measurement data, showing the results of observing vibration displacement of a specific device in the target system using a vibration sensor installed in the target system. The horizontal axis of the graph represents time, and this measurement data can be said to be measured in the time domain.

[0015] Once measurement data is obtained, the measurement data is transformed into a frequency domain representation, i.e., a frequency domain representation, to obtain a frequency spectrum. Fourier transform (FT) is commonly used as a method for transforming into a frequency domain representation, but other methods may also be used. The following describes the case where Fourier transform is used for the transformation into a frequency domain representation, but the results are the same even if other transformation methods are used to obtain a frequency spectrum. A fast Fourier transform (FFT) is preferably used as the Fourier transform method. FIG. 1(b) shows a frequency spectrum obtained by performing a Fourier transform on the measurement data shown in FIG. 1(a). In this graph, the horizontal axis represents frequency, and the vertical axis represents intensity, amplitude, power, etc. In the graph representing the frequency spectrum, the vertical axis may be expressed as an antilogarithmic number or a logarithmic number such as dB (decibels). In the method of this embodiment, one or more frequency bands are predetermined, and a partial spectrum is obtained by extracting portions belonging to each frequency band from a frequency spectrum obtained by a transformation such as a Fourier transform. If a frequency spectrum in the frequency range from 0 kHz to 15 kHz is obtained as a result of the Fourier transform, for example, as shown in FIG. 1(b), this frequency spectrum can be divided into four partial spectra: a partial spectrum in a frequency band B1 from 0 kHz to 2 kHz, a partial spectrum in a frequency band B2 from 2 kHz to 5 kHz, a partial spectrum in a frequency band B3 from 5 kHz to 10 kHz, and a partial spectrum in a frequency band B4 from 10 kHz to 15 kHz. The number of divisions when dividing the frequency spectrum into partial spectra is not limited to four and can be any number. In some cases, the frequency bands may partially overlap between the partial spectra. Of course, the entire frequency spectrum obtained by dividing the frequency spectrum into one may be treated as a single partial spectrum, or only one specific frequency band may be extracted from the frequency spectrum to obtain only one partial spectrum.

[0016] Once the partial spectrum is obtained, the coordinates of the center of gravity (centroid position) of the partial spectrum are calculated for each frequency band. Figure 1(c) shows an example of calculating the coordinates of the center of gravity position. The partial spectrum can be expressed as y = f(x) (where y ≥ 0) on the xy plane, with the horizontal axis (x-axis) being frequency and the vertical axis (y-axis) being intensity (or amplitude, power, etc.). Therefore, on the xy plane, the lower limit frequency of the frequency band corresponding to the partial spectrum is defined as f L , the upper frequency limit is f H Let x=f L , x=f H Consider the area enclosed by the three lines on the x-axis (i.e., y=0) and the curve expressed as y=f(x), and calculate the two-dimensional coordinate value G(x,y) of the center of gravity G of that area as a two-dimensional figure. In Figure 1(c), x=f L , x=f H The area surrounded by the x-axis and y=f(x) is hatched. The calculated two-dimensional coordinate value G(x, y) is used as a data vector to calculate the degree of abnormality, as will be described later.

[0017] As described above, one data vector used to calculate the degree of anomaly for each frequency band can be obtained from a single measurement data set from a sensor. In this embodiment, this process is repeated to obtain a series of multiple data vectors for each frequency band. The time required to obtain one measurement data set from a vibration sensor or acoustic sensor depends on the sampling frequency and number of sampling points when performing the Fourier transform, but is extremely short, for example, on the order of seconds or less. On the other hand, if the target system is, for example, a water treatment system, its operating time can extend from several months to several years. Therefore, to prevent an explosive increase in the number of data vectors and reduce the computational load, it is preferable to set an appropriate processing unit period and calculate one data vector for each appropriate processing unit period for each frequency band. In this case, measurement data can be obtained by performing one measurement within one processing unit period, or by performing multiple measurements, averaging the measurement data, and performing a Fourier transform on the averaged measurement data to calculate a data vector. Furthermore, multiple measurements can be performed, a Fourier transform can be performed for each measurement, and the results obtained by the Fourier transform can be averaged to calculate a data vector.

[0018] The processing unit period can be determined appropriately depending on the operating mode of the target system. For example, if the target system is equipped with an ion exchange device that alternately performs a water flow process and a regeneration process, one water flow process of the ion exchange device can be set as one processing unit period. Of course, one water flow process can also be divided into multiple measurement periods. The lengths of the processing unit periods may differ from one another. As another example, the processing unit period can be set at regular intervals, for example, every 10 minutes or every hour, while the target system is operating.

[0019] Here, the elements of a data vector are described as two-dimensional coordinate values ​​of the center of gravity of a partial spectrum in a specific frequency band extracted from measurement data from a single sensor. Therefore, the data vector is a two-dimensional vector. However, the data vector may include other data obtained within the same processing unit period as elements, in which case the number of dimensions of the data vector is greater than two. For example, when L (L≧2) partial spectra in different frequency bands are obtained from measurement data from sensors, the coordinates of the center of gravity for each of those L partial spectra can be calculated for each processing unit period, and the obtained two-dimensional coordinate values ​​can be combined to form a single data vector. In this case, the data vector is a 2L-dimensional vector. Alternatively, when multiple sensors are provided, measurement data can be acquired from each sensor, partial spectra can be calculated in the same manner as above to determine the coordinates of the center of gravity, and the two-dimensional coordinate values ​​of the center of gravity obtained for each sensor can be combined to form a data vector. In this case, the number of dimensions of the data vector is also greater than two. Data other than vibration and acoustic data obtained within a processing unit period may also be included as elements in the data vector. The data vector calculated in this manner is stored in a memory or the like.

[0020] Next, the degree of anomaly of each data vector is calculated based on the accumulated data vectors. For example, the Mahalanobis distance MD can be used as the degree of anomaly, but other indices can also be used as the degree of anomaly. For example, a value obtained by performing a singular spectrum transform on the data vector can also be used as the degree of anomaly. In the following, detection of a sign of an anomaly occurrence in this embodiment will be described assuming that the Mahalanobis distance MD is used as the degree of anomaly.

[0021] Here, the calculation of the Mahalanobis distance MD will be explained. If a series of processing unit periods are assigned sequential numbers starting from 1, then the N-dimensional data vector a in processing unit period i is iis given by equation (1). As mentioned above, when a data vector is constructed using only the two-dimensional coordinate values ​​G(x, y) of the center of gravity G of the partial spectrum of a specific frequency band, N=2. Also, as mentioned above, if the two-dimensional coordinate values ​​of the center of gravity obtained from the partial spectra of L different frequency bands are combined into one data vector, the data vector is a vector formed by arranging L two-dimensional vectors, and in this case, the data vector is a 2L-dimensional vector (i.e., N=2L). Specifically, when a data vector is expressed as in equation (1), a i1 ,a i2 are the x- and y-coordinate values ​​of the center of gravity obtained from the first partial spectrum in the processing unit period i, and a i3 ,a i4 are the x- and y-coordinate values ​​of the center of gravity obtained from the second partial spectrum in the processing unit period i, respectively, and a i(N-1) ,a iN are the x- and y-coordinate values ​​of the center of gravity obtained from the L-th partial spectrum in the i-th processing unit period. T represents the transpose. a i =(a i1 ,a i2 ,…,a iN ) T (1)

[0022] If μ is an N-dimensional vector indicating the average value of data vector a over k processing unit periods, μ is given by equation (2). μ = (μ1, μ2, …, μ N ) T (2) however, μ j =(a 1j +a 2j +…+a kj ) / k

[0023] If the covariance matrix calculated from the k stored data vectors is Σ, then the data vector a obtained in the processing unit period i is i The Mahalanobis distance MD i is given by equation (3).

[0024]

number

[0025] A large Mahalanobis distance for a given processing unit period means that the data from that processing unit period tends to deviate from the data from many other processing unit periods. Figures 1(d) and 1(e) show the distribution of a series of two-dimensional coordinate values ​​of the center of gravity (i.e., two-dimensional data vectors) calculated from partial spectra obtained by sampling measurement data from the same sensor for each repeated processing unit period and performing a Fourier transform. The frequency bands used to extract partial spectra from the frequency spectrum obtained by the Fourier transform are different between Figures 1(d) and 1(e). The group of ellipses in Figures 1(d) and 1(e) are contour lines (equal Mahalanobis distance lines) for the Mahalanobis distance MD. The Mahalanobis distance increases as you move away from the center of the group of ellipses. The group of ellipses may have shapes that are close to circles (ellipses with small eccentricity), shapes that are elongated in the y-axis direction (vertically elongated shapes such as the shape shown in Figure 1(d)), shapes that are elongated in the x-axis direction (horizontally elongated shapes such as the shape shown in Figure 1(e)), and shapes that are elongated in a diagonal direction, i.e., in a direction that is not parallel to either the x-axis or the y-axis (elongated shapes that rise to the right or left).

[0026] Once the Mahalanobis distance is calculated for the data vector for each processing unit period, the Mahalanobis distance is compared with a predetermined threshold value, for example. If the Mahalanobis distance is greater than the threshold value, it can be determined that there is a sign of an abnormality occurring in that processing unit period. Furthermore, in this embodiment, data vectors for new processing unit periods are calculated and accumulated over time. Therefore, when the Mahalanobis distance is recalculated including the new data vectors, the Mahalanobis distance of the newly generated data vectors may be significantly greater than the Mahalanobis distance of the previously generated data vectors. This strongly suggests that there is a sign of a new abnormality occurring. For example, as described above, the group of ellipses representing equal Mahalanobis distance lines can have various shapes. However, if the measurement data is vibration measurement data and the group of ellipses is elongated in the y-axis direction or diagonally, it can be said that this indicates that the vibration amplitude and intensity are large and that the variations in these vibrations are large. Therefore, it can be determined that there is a sign of an abnormality occurring. Furthermore, if we take the center of the group of ellipses as a reference point and draw a straight line (horizontal line) parallel to the x-axis so that it passes through this reference point, when the position of the data vector spreads above the horizontal line over time, it means that vibrations and amplitudes are increasing over time, and we can conclude that this is a sign of an abnormality. Conversely, when the position of the data vector spreads below the horizontal line over time, it means that vibrations and amplitudes are decreasing over time, and we can conclude that the possibility of an abnormality occurring is low.

[0027] If the sensor is a vibration or acoustic sensor, the processing unit period can be set short, allowing for multiple processing unit periods and the acquisition of multiple data vectors within a period during which the target system is known to be operating normally. Therefore, the data vectors acquired when the target system is operating normally can be set as a reference data set. Subsequently, the Mahalanobis distance from the reference data set can be calculated for each new data vector acquired. Figure 1(f) shows an example of how the Mahalanobis distance from the reference data set for data vectors acquired in each processing unit period changes over the course of the processing unit period. The black circles and open squares in the figure represent the Mahalanobis distances calculated from data vectors acquired from partial spectra in different frequency bands, respectively. If the calculated Mahalanobis distance MD increases over time, it can be determined that there are signs of an abnormality.

[0028] FIG. 2 is a flowchart illustrating an anomaly sign detection method according to one embodiment of the present invention. Here, a case will be described in which partial spectra are extracted from a frequency spectrum obtained from measurement data from one sensor and a Mahalanobis distance is calculated. First, in step 101, measurement data from a sensor in a target system is sampled and a Fourier transform is performed to obtain a frequency spectrum. In step 102, one or more frequency bands are set for the frequency spectrum, and a partial spectrum is obtained for each frequency band. In step 103, two-dimensional coordinate values ​​of the centers of gravity (centroids) of the partial spectra are calculated as data vectors, and the data vectors are stored. Then, in step 104, it is determined whether the number of data vectors already stored and accumulated is sufficient for calculating the Mahalanobis distance. If not, the process from step 101 is repeated to continue accumulating data vectors. On the other hand, if a sufficient number of data vectors have been accumulated, in step 105, the Mahalanobis distance is calculated for each accumulated data vector, and in step 106, a sign of an abnormality is detected based on the calculated Mahalanobis distance. Thereafter, in step 107, it is determined whether or not to continue detecting signs of an abnormality. If it is determined to continue detecting signs, in step 108, measurement data for the next processing unit period is acquired and a Fourier transform is performed, and in step 109, a data vector corresponding to the next processing unit period is generated and stored, and then the processing from step 105 is repeated.

[0029] Execution of steps 108 and 109 increases the number of accumulated data vectors. However, when step 105 is executed after execution of step 109, the Mahalanobis distance is recalculated for all accumulated data vectors, including the increased data vectors. Alternatively, the number M of data vectors used to calculate the Mahalanobis distance may be determined in advance, and when step 105 is executed, the Mahalanobis distance may be calculated using the most recent M data vectors among the accumulated data vectors. M may be a value based on the duration of one operation of the target system, for example, equivalent to several operating times. On the other hand, when it is determined in step 107 that the detection of signs of abnormality will not continue, the process of detecting signs of abnormality is terminated.

[0030] While manual analysis of vibration data and acoustic data to detect signs of abnormality requires skilled techniques, the method of this embodiment automatically calculates the Mahalanobis distance and the contribution, making it possible to easily detect signs of abnormality in the target system and analyze the cause of the abnormality. The method of this embodiment also has the advantage that it can be executed without using data obtained when the target system is operating normally (a reference data set), and can be easily applied to predictive analysis using techniques such as machine learning.

[0031] FIG. 3 shows an example of the configuration of an abnormality sign detection device 10 that implements the abnormality sign detection method described above. This abnormality sign detection device 10 detects signs of abnormality in a target system 20 in order to manage the operation of the target system 20. The target system 20 is provided with a sensor 21, such as a vibration sensor or an acoustic sensor. Examples of vibration sensors used as the sensor 21 include a MEMS (micro-electromechanical system) sensor, a piezoelectric sensor, and a laser Doppler sensor, and examples of acoustic sensors include a microphone. A plurality of sensors 21 may be provided. Furthermore, the sensor 21 may measure voltage, current, power, etc., and may be a power analyzer, etc. When measuring voltage or current to detect signs of abnormality, an oscilloscope or the like may also be used as the sensor 21. Measurement data from each sensor 21 is input to the abnormality sign detection device 10.

[0032] The anomaly sign detection device 10 includes a conversion calculation unit 11, which is provided for each sensor 21 and samples measurement data sent from the sensor 21, converts the data into a frequency domain representation, and acquires a frequency spectrum for each processing unit period; a centroid calculation unit 12, which extracts a portion of the frequency spectrum for each frequency band based on one or more predetermined frequency bands to generate a partial spectrum and calculates the two-dimensional coordinate value of the centroid of the partial spectrum; a data storage unit 13, which accumulates data vectors consisting of the two-dimensional coordinate values ​​calculated for each processing unit period by the centroid calculation unit 12; and an anomaly degree calculation unit 14, which calculates the anomaly degree for each series of data vectors stored in the data storage unit 13. The conversion calculation unit 11 acquires the frequency spectrum by, for example, performing an FFT operation on the measurement data. The anomaly degree calculation unit 14 calculates, for example, the Mahalanobis distance MD as the anomaly degree. When a data vector has a high anomaly degree, an administrator operating the target system 20 can determine that there was a sign of an anomaly in the target system 20 during the processing unit period corresponding to that data vector.

[0033] The abnormality sign detection device 10 may include a determination unit 15 that determines whether or not there are signs of an abnormality in the target system 20 based on the calculated abnormality degree, and, if there are signs of an abnormality, estimates the cause of the abnormality. The determination unit 15 outputs a determination result indicating whether or not an abnormality has occurred and the estimated cause of the abnormality. Furthermore, by incorporating a machine learning engine that performs machine learning into the determination unit 15, it is possible to improve the accuracy of detecting signs of an abnormality and identifying the location of the abnormality.

[0034] Next, an example will be described in which the above-described abnormality sign detection method is applied to an actual system. Here, the target system 20 is an ultrapure water production system 100 that is supplied with primary pure water and produces ultrapure water. Figure 4 shows an example of the configuration of the ultrapure water production system 100.

[0035] The ultrapure water producing system 100 shown in FIG. 4 includes a primary pure water tank 110 that receives primary pure water supplied from the primary pure water system. The water in the primary pure water tank 110 is treated as water to be treated, and the water is sequentially treated to produce ultrapure water, which is then supplied to a point-of-use 170. Any ultrapure water not consumed at the point-of-use 170 is returned to the primary pure water tank 110 via a circulation pipe 175. A pump 120 is provided at the outlet of the primary pure water tank 110 to pressurize and feed the water in the primary pure water tank, i.e., the water to be treated. A device group 130, consisting of a combination of devices such as a heat exchanger, an ultraviolet irradiation device, and an ion exchange resin tower, is provided on the secondary side of the pump 120. The outlet water from the device group 130 is supplied to a membrane degassing device (MD) 140. Nitrogen (N) gas is supplied to the membrane degassing device 140, and a vacuum pump (VP) 145 is connected to the membrane degassing device 140. Two boost pumps 150A and 150B are provided in parallel at the outlet of the membrane degassing device 140 to boost the pressure of the outlet water of the membrane degassing device 140. The secondary sides of the boost pumps 150A and 150B join together, where an equipment group 160 is provided that combines devices such as an ion exchange resin tower and an ultrafiltration membrane device. The outlet water from the equipment group 160 is supplied as ultrapure water to a point of use 170 via a supply pipe 165.

[0036] If the pressure loss between the ultrapure water producing system 100 and the point of use is large, including due to the height of the point of use, it is necessary to increase the pressure at the outlet side of the equipment group 160. Increasing the discharge pressure of the pump 120 for this purpose could adversely affect the membrane degassing device 140. The boost pumps 150A and 150B are provided to increase the delivery pressure at the outlet of the equipment group 160 without adversely affecting the membrane degassing device 140. In this ultrapure water producing system 100, the vacuum pump 145, the boost pump 150A, and the supply pipe 165 are each provided with a vibration sensor 21 that detects vibrations. Measurement data obtained by these sensors 21 is sent to the abnormality sign detection device 10, as shown in FIG. 3 . Instead of a vibration sensor, an acoustic sensor may be used as the sensor 21 attached to the vacuum pump 145 or the boost pump 150A, or a sensor that measures the voltage, current, or voltage of the motor that drives the pump may be used.

[0037] The ultrapure water production system 100 shown in FIG. 4 was operated for approximately 16 months, and the abnormality sign detection method of this embodiment was applied. The Mahalanobis distance was calculated as the degree of abnormality, and the change in the Mahalanobis distance over time was examined. Here, a case where the Mahalanobis distance was calculated for each frequency band based on measurement data from the sensor 21 attached to the boost pump 150A is described. The results are shown in FIG. 5. Note that data from periods when the operation of the ultrapure water production system 100 was temporarily suspended is excluded from FIG. 5. Furthermore, at the start of operation, the two boost pumps 150A and 150B were operated in parallel. However, at time P, the boost pump 150B abnormally stopped, and thereafter, only the boost pump 150A was operated. Then, at time Q, the boost pump 150B recovered, and thereafter the two boost pumps 150A and 150B resumed operating in parallel.

[0038] In Figure 5, (a) shows the change in Mahalanobis distance calculated from the partial spectrum obtained in the frequency band of 0 to 15 kHz. Similarly, (b) to (e) show the change in Mahalanobis distance calculated from the partial spectrum obtained in the frequency bands of 0 to 2 kHz, 2 to 5 kHz, 5 to 10 kHz, and 10 to 15 kHz, respectively. As can be seen from Figure 5, in the measurement data from sensor 21 attached to boost pump 150A, the abnormal shutdown of the other boost pump 150B is represented by a large change in the Mahalanobis distance calculated from the center of gravity of the partial spectrum in the frequency band of 10 to 15 kHz. In contrast, the Mahalanobis distances calculated from the partial spectra obtained in the frequency bands of 0 to 2 kHz, 2 to 5 kHz, and 5 to 10 kHz show almost no effect of the abnormal shutdown of boost pump 150B. On the other hand, the effect of the recovery of boost pump 150B is reflected in the Mahalanobis distances calculated from the partial spectra obtained in all frequency bands. In particular, large changes are observed in the Mahalanobis distance calculated from the partial spectrum obtained from the frequency band of 2 to 5 kHz.

[0039] From the above, it can be seen that by setting one or more frequency bands to the frequency spectrum obtained from the measurement data by Fourier transform, extracting a partial spectrum for each frequency band, determining the two-dimensional coordinate value of the center of gravity of each partial spectrum for each processing unit period and accumulating them as a data vector, and calculating the Mahalanobis distance based on the accumulated data vector, it is possible to detect signs of an abnormality occurring in the target system and also to estimate the location and content of the abnormality. [Explanation of symbols]

[0040] 10. Anomaly detection device 11 Conversion calculation unit 12 Center of gravity calculation section 13 Data storage unit 14 Abnormality calculation unit 15 Judgment section 20 Target System 21 Sensors 100 Ultrapure water production system 110 Primary pure water tank 120 Pump 130,160 equipment group 140 Membrane degassing device 150A, 150B Booster Pump 170 Use Points

Claims

1. An abnormality sign detection method for detecting a sign of an abnormality occurring in a target system, comprising: a spectrum acquisition step of converting measurement data obtained from the target system into a frequency domain representation to acquire a frequency spectrum; a centroid calculation step of extracting a portion belonging to each of the predetermined frequency bands from the frequency spectrum to obtain a partial spectrum based on one or more predetermined frequency bands, determining coordinate values ​​of the centroid of the partial spectrum, and calculating a data vector including at least the coordinate values ​​as elements; and the spectrum acquisition step and the centroid calculation step are repeatedly performed to accumulate a series of the data vectors, and the degree of abnormality of each of the data vectors is calculated.

2. The abnormality sign detection method according to claim 1 , further comprising detecting a sign of an abnormality occurring in the target system based on the calculated abnormality degree.

3. The abnormality sign detection method according to claim 1 , wherein the spectrum acquisition step and the centroid calculation step are performed for each of a series of processing unit periods defined in the target system.

4. Repeating the calculation and accumulation of the data vector for new processing unit periods that occur over time; 4. The abnormality sign detection method according to claim 3, further comprising detecting a sign of an abnormality occurring in the target system based on a change in the abnormality degree corresponding to the new processing unit period added over time.

5. The abnormality sign detection method according to claim 1 , wherein the measurement data is measurement data relating to sound or vibration.

6. The abnormality sign detection method according to claim 1 , wherein the degree of abnormality is a Mahalanobis distance.

7. 5. The abnormality precursor detection method according to claim 1, wherein the target system is a water treatment system including a pump, and the measurement data is measurement data from at least one of a vibration sensor, an acoustic sensor, a voltage sensor, a current sensor, and a power sensor attached to the target system.

8. An abnormality sign detection device that detects a sign of an abnormality occurring in a target system, a conversion calculation unit that converts measurement data obtained from the target system into a frequency domain representation to obtain a frequency spectrum; a centroid calculation unit that extracts a portion belonging to each of the predetermined frequency bands from the frequency spectrum for each of the predetermined frequency bands, and calculates a coordinate value of the centroid of the partial spectrum; a data storage unit for storing a data vector including the coordinate values ​​calculated by the center of gravity calculation unit as elements; an abnormality degree calculation unit that calculates an abnormality degree of the data vector based on the series of data vectors stored in the data storage unit; An abnormality sign detection device having the above.

9. The abnormality sign detection device according to claim 8 , further comprising a determination unit that determines whether or not there is a sign of an abnormality occurring in the target system based on the abnormality degree.

10. 10. The abnormality sign detection device according to claim 8, wherein the measurement data is measurement data from at least one of a vibration sensor, an acoustic sensor, a voltage sensor, a current sensor, and a power sensor attached to the target system.

Images