JP7806507B2 - Equipment diagnosis device and equipment diagnosis method - Google Patents

Description

The present invention relates to technology for diagnosing equipment based on measurement data from multiple sensors installed in the equipment to be diagnosed.

Failures in infrastructure facilities that support society, such as those for electricity and transportation, can have a significant impact on society, making it necessary to prevent unplanned shutdowns before they occur. Furthermore, failures in production equipment in factories can also have a significant impact on a company's production activities.

As these facilities age, there is a growing need to reduce costs amid a labor shortage due to an aging population and declining birthrate, as well as increased global competition. Therefore, there is a growing demand for unmanned, remote monitoring systems that can detect equipment failures and abnormalities early on.

In order to monitor equipment 24 hours a day, 356 days a year, diagnostic technology has been proposed that involves attaching many sensors to the equipment to be monitored and using the multidimensional time series signals detected by each sensor.

For example, Patent Document 1 proposes measuring the vibrations of a rotating machine using a vibration sensor and diagnosing the results of the measurement. Furthermore, Patent Document 2 achieves diagnosis by measuring the current flowing through the grounding wire. According to these diagnoses, signs of an abnormality appear in vibrations and currents, and measuring these makes it possible to detect an abnormality.

Patent Document 3 proposes that increasing the number of state quantities (physical quantities) used to determine abnormalities will enable more accurate abnormality detection and failure prediction.

Patent Publication No. 2021-144054 Patent Publication No. 2021-50921 Patent Publication No. 2021-177074

"Introduction to the world of divergence (KL-divergence) and entropy (cross-entropy) used in machine learning" [online] Retrieved December 27, 2021, Internet <URL: https://qiita.com/harmegiddo/items/2a24a36418fade0eaf44> Shota Ako, "Machine Learning and Differential Geometry", Mathematical Sciences, Vol. 56, no. 8, pp. 14-19, Aug. 2018 S. Kullback and R. A. Leibler,”On information and sufficiency Ann. Math. Statist”, PP. :79-,1951 S.Amari and H.Nagaoka, “Methods of Information Geometry” A co-publication of the AMS and Oxford University Press, 2000. Takuya Ajiki, Masami Nakamura, Shunji Maeda, Takashi Komatsu, Tetsuji Taniguchi, "Proposal of an Anomaly Detection Method Using KLD Duality", 72nd Chugoku Branch Joint Conference of Electrical and Information Engineering Societies, 2021

However, when using data from multiple sensors, if the physical quantities measured are different, the characteristics of the data (such as units and statistics) will change, making it difficult to understand the relationships between them and potentially making correlation analysis difficult to perform.

The present invention aims to simplify diagnosis using multiple sensor data by performing diagnosis based on an evaluation of the correlation between sensor data.

(1) One aspect of the present invention is
A device that installs a plurality of sensors in a facility to be diagnosed, and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data,
a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each sensor, divides the set distribution interval into a number of classes that indicate a preset random variable, and converts the set distribution interval into a probability distribution according to the number of measurement values of each class;
a mutual information conversion unit that calculates mutual information for the same time from the probability distribution of each sensor converted by the probability distribution conversion unit using equation (3);

P _X (x): Probability distribution of random variable X = {x ₁ , x ₂ , . . . x _n } of one sensor P _Y (y): Probability distribution of random variable Y = {y ₁ , y ₂ , . . . _yn } of the other sensor where random variable X = random variable Y
I(X;Y): Mutual information of P _X (x) and P _Y (y) P _XY (x, y): Joint probability (relative frequency when data from one sensor and data from another sensor are included in each class at the same time)
Equipped with
The equipment is diagnosed based on a correlation change in the mutual information calculated by the mutual information transform unit.

(2) Another aspect of the present invention is a method executed by a device that installs a plurality of sensors in a facility to be diagnosed, and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data, the method comprising:
a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each sensor, dividing the set distribution interval into a number of classes representing a preset random variable, and converting the set distribution interval into a probability distribution according to the number of measurement values of each class;
a mutual information conversion step of calculating mutual information for the same time from the probability distribution of each sensor converted in the probability distribution conversion step using equation (3);

P _X (x): Probability distribution of random variable X = {x ₁ , x ₂ , . . . x _n } of one sensor P _Y (y): Probability distribution of random variable Y = {y ₁ , y ₂ , . . . _yn } of the other sensor where random variable X = random variable Y
I(X;Y): Mutual information of P _X (x) and P _Y (y) P _XY (x, y): Joint probability (relative frequency when data from one sensor and data from another sensor are included in each class at the same time)
a diagnosis step of diagnosing the equipment based on a correlation change of the mutual information calculated by the mutual information transform unit;
It is characterized by having:

(3) In each of the above aspects,
generating a reference probability distribution P _X ^* (x) from the probability distribution;
A probability distribution P _Y (x) to be compared with the same random variable " _X ⁼ Y = {x 1 , x _{2 ,} . . . x _{n }} " as the probability distribution P _X ^* (x) is obtained from the probability distributions P _X (x) and P Y (y) of each sensor obtained at the same intervals as the probability distribution P _{X *} (x);
The probability distributions P _X ^* (x) and P _Y (x) are converted into the KLD "Kullback-Leibler divergence/D _KL (X∥Y)" of equation (4),

The equipment can also be diagnosed based on whether the transformed KLD is detected as an outlier away from the normal distribution in a scatter plot.

At this time, the probability distribution P _X ^* (x) and the P _Y (x) are interchanged and converted into the KLD "D _KL (Y∥X)" of equation (5),

The equipment may be diagnosed based on whether the KLDs of equations (4) and (5) are both detected as outliers away from the normal distribution in the scatter plot.

(4) In addition, in each of the above aspects,
The probability distribution of each sensor converted by the probability distribution conversion unit is converted into an entropy for each arbitrary time using equations (1) and (2);

H(X): Entropy of one sensor H(Y): Entropy of the other sensor The equipment can also be diagnosed based on whether or not the entropy of each sensor has changed by a threshold or more.

This invention makes it possible to facilitate diagnosis using data from multiple sensors.

1A is a schematic diagram of the diagnostic information amount (entropy) of an equipment diagnostic system according to an embodiment of the present invention, FIG. 1B is a schematic diagram of mutual information, and FIG. 1C is a schematic diagram of "KLD". FIG. 2 is an overall diagram showing the processing steps of the diagnostic method. FIG. 2 is a diagram showing a diagnostic status of a target facility according to an embodiment. FIG. 2 is an overall diagram showing the processing steps of the diagnostic method. 1A is a diagram of sensor data for the sensor 12, and FIG. 1B is a diagram of sensor data for the sensor 13. FIG. 1A is a diagram of time-series sensor data from the sensor 12 input to the probability distribution conversion unit, and FIG. 1B is a diagram of data output from the probability distribution conversion unit after the time-series sensor data in FIG. 1A is converted into a probability distribution. 1A is a data diagram showing the probability distribution of the sensor 12 at regular intervals input to the entropy conversion unit, and FIG. 1B is a data diagram output from the entropy conversion unit after the probability distribution of FIG. 1A is converted into entropy. (a) is a data diagram showing the probability distribution of sensor 12 input to the mutual information conversion unit, (b) is a data diagram showing the probability distribution of sensor 13, and (c) is a data diagram output from the mutual information conversion unit after the probability distributions of (a) and (b) are converted into mutual information. (a) is a data diagram of the reference probability distribution (sensor 12) input to the KLD conversion unit, (b) is a data diagram of the probability distribution (sensor 13) to be compared, and (c) is an outlier detection diagram of the output results of the KLD conversion unit obtained from the probability distributions of (a) and (b).

The following describes an equipment diagnostic device (equipment diagnostic method) according to an embodiment of the present invention. This diagnostic device converts the phenomenon being diagnosed into numerical values (data) as physical quantities, and expresses these numerical values as "statistics such as averages and variances." This is then converted into information value, and the "rarity" of the information is used as a diagnostic index.

Specifically, the diagnostic device treats sensor data with various units from multiple sensors attached to the equipment being diagnosed as a probability distribution based on occurrence frequency, and focuses on the entropy and mutual information calculated from each sensor data to diagnose the presence or absence of abnormalities or signs of abnormalities in the equipment being diagnosed.

Evaluation methods based on such multiple sensor data include "JSD (Jensen-Shannon Divergence)" in Non-Patent Document 1, "Wasserstein" in Non-Patent Document 2, and "KLD (Kullback-Leibler Divergence)" in Non-Patent Documents 3-5.

In response to this, the diagnostic device combines three information quantities (entropy, mutual information, and KLD) to detect abnormalities and signs of abnormalities that could not be detected by "JSD," "Wasserstein," or "KLD" alone.

When using multiple sensor data, each sensor data changes in a variety of ways and uses different units, making it difficult to handle. Therefore, the diagnostic device converts sensor data of different physical quantities collected from various sensors attached to the equipment being diagnosed into a probability distribution based on occurrence frequency, and then converts it into a common unit such as entropy or mutual information, which indicates the amount of information. This unifies the unit system, and furthermore, evaluates the correlation between the sensor data under the unified unit system, making it easier to diagnose using multiple sensor data.

(1) Amount of Information Diagnosis based on each amount of information will be outlined with reference to FIG.

A: As shown in Figure 1(a), the entropy (average amount of information) H of each sensor data is calculated at regular intervals. At this time, abnormalities in the state of each sensor data are detected based on the temporal changes in each calculated entropy H, and the presence or absence of abnormalities/signs of abnormalities in the diagnostic target is diagnosed. Here, equipment abnormalities and the like are considered as changes in entropy H, thereby unifying the unit system.

B: As shown in Figure 1(b), the mutual information I between sensor data is calculated at regular intervals. Sensor data showing abnormal conditions is detected from the correlation based on the calculated mutual information I, and the presence or absence of an abnormality or signs of an abnormality in the target is diagnosed. In other words, by calculating the mutual information between sensor data, the equipment is diagnosed based on the dependency between sensor data under normal conditions and the temporal changes in the relationship when changing from normal to abnormal. This diagnosis is not affected by physical quantities, so detection is stable.

C: As shown in Figure 1(c), the "KLD" between a pre-recorded reference probability distribution (reference data) and the probability distribution of the object to be diagnosed is calculated at regular intervals. Here, abnormalities in the sensor data are detected from changes in the "KLD," and the presence or absence of an abnormality/sign of an abnormality in the object to be diagnosed is diagnosed. In this case, the reference probability distribution and the probability distribution of the object to be compared do not need to be sensor data from the same sensor.

(2) Diagnostic Procedure The overall diagnostic procedure (diagnostic steps) performed by the diagnostic device will be outlined with reference to Fig. 2. First, information that quantifies the physical phenomenon of the diagnostic target, i.e., sensor data from the sensors 1 and 2 of the diagnostic target, is acquired (S01).

Next, in order to express each acquired sensor data as a statistical index, each sensor data is converted into a probability distribution " _PX (x), _PY (y)" (S02). The converted probability distribution " _PX (x), _PY (y)" is expressed as an index of "information value."

Here, the probability distribution "P _X (x), P _Y (y)" is defined as
・Entropy H(X), H(Y)
・Mutual information I(X;Y)
・KLD “Bivariate D _KL (X||Y), D _KL (Y||X)”
(S03 to S05). After that, the numerical values of the information converted in S03 to S05 are monitored (S06 to S08).

That is, in S06, "Entropy H Monitoring (Self-Status Diagnosis)," the increase or decrease in the entropy H value is monitored. As a result of the monitoring, it is determined that "increase in value = increased variability" and "decrease in value = increased numerical deviation," and it is determined that an abnormality or a sign of an abnormality has occurred in the object being diagnosed.

Furthermore, in S07, "Monitoring the value of mutual information I," the system determines whether the relationship between sensors 1 and 2 has been disrupted by determining whether the mutual information I is "increasing in value = strong relationship between sensors 1 and 2" or "decreasing in value = weak relationship (independence)." If the monitoring results in a disruption of this relationship, it is determined that an abnormality or a sign of an abnormality has occurred in the object to be diagnosed.

Furthermore, in S08, "KLD Numerical Value Monitoring," it is monitored whether a change in distribution has occurred in the two variables. If a change in distribution is found as a result of the monitoring, it is determined that an abnormality or a sign of an abnormality has occurred in the object to be diagnosed.

An example of the diagnostic device will be described with reference to Figures 3 to 9. Here, a rotating machine will be used as an example of equipment.

<Configuration example>
3, reference numeral 11 denotes a rotating machine that is the equipment to be diagnosed. An acceleration sensor 12 is installed near the bearing of this rotating machine 11 to measure bearing vibration (acceleration).

In addition, a current sensor 13 is installed in the housing 11a of the rotating machine 11 to measure the power supply current and leakage current of the rotating machine 11. The measurement data (sensor data) measured by these two sensors 12, 13 is collected and stored in a data measuring device 14 equipped with a storage device (e.g., SSD, HDD, etc.).

Here, an equipment diagnosis device 10 is connected to the data measuring instrument 14, and the collected and accumulated measurement data is input to the equipment diagnosis device 10. This equipment diagnosis device 10 generates a probability distribution based on the input measurement data and diagnoses the rotating machine 11 using the various information quantities (entropy, mutual information, KLD). The diagnosis results can be displayed on a display device (monitor) (not shown) connected to the diagnosis device 10, allowing the user to check them.

Specifically, the diagnostic device 10 is configured as a computer and is equipped with the usual hardware resources of a computer (e.g., a CPU, RAM/ROM storage devices, etc.).

As a result of the collaboration between these hardware resources and software resources (OS, applications, etc.), the diagnostic device 10 implements a probability distribution transformation unit, entropy transformation unit, mutual information transformation unit, KLD transformation unit, and monitoring unit (diagnosis unit), all of which are not shown.

<Diagnostic Procedure>
The diagnostic procedure (diagnostic steps) of the embodiment will be described with reference to FIG.

S11: Sensors 12 and 13 continuously measure physical information (acceleration and current value) of the rotating machine 11 being diagnosed. The sensor data measured here is input into the data measuring device 14.

At this time, the data measuring device 14 periodically records each captured sensor data as time-series data of an arbitrary time width. Figure 5 shows an example of time-series data from sensors 12 and 13, recorded with one measurement per second. Here, the sampling frequency of the time-series data from sensors 12 and 13 is 51,200 Hz, and 51,200 samples of data are recorded per second.

As a result, diagnostic time series data is accumulated every second in the data measuring device 14, and the accumulated time series data is output to the diagnostic device 10. Note that D1 in Figure 5 indicates an abnormality (100 Hz contamination) in the time series data (4.0 to 5.0) of the sensor 12.

S12: The time series data from the sensors 12 and 13 accumulated in the data measuring device 14 is input to the diagnostic device 10. Based on the time series data input here, the probability distribution conversion unit generates (calculates) a probability distribution P for each fixed time period. At this time, the following parameters are prepared:

(1) Distribution interval [a, b]
The distribution interval [a, b] indicates the value that determines the range of the probability distribution. It is desirable to set this range wider than the range from the maximum value to the minimum value of the measurement value. For example, in the time-series sensor data of Figure 6(a), the measurement value is included in "±0.3", and the distribution interval [a, b] = [-0.4, 0.4] (±0.4) is set.

(2) Number of classes "n"
The number of classes "n" indicates a value that determines the resolution when dividing the measurement values into several sections.

The distribution interval [a, b] is divided into subintervals (hereafter referred to as "classes") by the number of classes "n". For example, if the distribution interval is [-0.4, 0.4] and the number of classes "n" = 40, the series of classes will be 40 divisions: {[-0.4, -0.38], [-0.38, -0.36], ..., [+0.38, +0.36].

Then, the number of measurement values included in each class (hereinafter referred to as frequency) is calculated. Taking the measurement values of the sensor 12 as an example, the classes are set as "X = { _x1 , _x2 , ..., _xn }" and are treated as random variables. In this case, the probability distribution P _X of the sensor 12 is expressed as the value obtained by dividing the frequency of each class by the number of sample data.

For example, if the frequency sequence is {0, 0, 0, 1, 6, 15, 33, 72, ..., 0} and the number of sample data is "51200", then P _X (x) = {0, 0, 0, 0, 0.0001, 0.0003, 0.0006, 0.0014, 0.0031, 0.0055, ..., 0}, and the probability distribution P _X (x) shown in D2 in Figure 6(b) is obtained.

A similar process is performed on the measurement values of the sensor 13 to obtain the probability distribution P _Y (y) of the sensor 13. The process of S12 is performed on the sensor data of each of the sensors 12 and 13 every time the process of S11 is performed.

As described above, since sensor data is collected with one measurement per second, the latest probability distributions P _X (x) and P _Y (y) are calculated every second. The calculated probability distributions P _X (x) and P _Y (y) are stored in a storage device (HDD, SSD, etc.) and are output to the entropy conversion unit, mutual information conversion unit, and KLD conversion unit.

S13: The entropy conversion unit calculates the entropy H for the sensor data of each of the sensors 12 and 13 at regular intervals every time the probability distributions P _X (x) and P _Y (y) are input.

When the entropy H(X) is calculated for the probability distribution P _X (x) of the sensor 12 shown in Fig. 7(a), the output result shown in Fig. 7(b) is obtained. In this case, if the probability distribution consisting of the random variable X = {x ₁ , x ₂ , ..., x _n } is defined as P _X (x), the entropy H(X) can be found from equation (1).

Similarly, the entropy H(Y) is calculated for the probability distribution P _Y (y) of the sensor 13. In this case, if the probability distribution consisting of the random variable Y={y ₁ , y ₂ , . . . , _yn } is defined as P _Y (y), the entropy H(Y) can be obtained from equation (2).

However, entropy H is calculated using the same random variable in the sensor data from sensors 12 and 13 (random variable X = random variable Y). In this respect, it is possible to unify the unit system.

S14: The monitoring unit monitors changes in the entropies H(X) and H(Y) calculated in S13, and diagnoses whether there is an abnormality or a sign of an abnormality each time entropy H(X) and H(Y) are calculated. This diagnosis is performed for each entropy H(X) and H(Y) based on a preset threshold. This threshold is determined based on an empirically obtained range of change.

D3 (entropy H(X) after 5 seconds) in Figure 7(b) shows a state in which the entropy H(X) calculated from the probability distribution D2 exceeds the threshold. In this case, entropy H(X) increases after 4 seconds, and after 5 seconds it exceeds the threshold, indicating an abnormal value. In this case, the rotating machine 11 is diagnosed as having an abnormality/sign of an abnormality, and the diagnosis result is displayed on the display device. Note that the threshold can also be set in stages for each abnormality/sign of an abnormality, and each abnormality/sign of an abnormality can be displayed on the display device.

S15: The mutual information converter calculates the mutual information I for the same time from the probability distributions P _X (x) and P _Y (y) every time the probability distributions P _X (x) and P _Y (y) of S12 are input. For example, if the mutual information I per second is calculated from the probability distributions P _X (x) and P _Y (y) of Figures 8(a) and 8(b), the output result shown in Figure 8(c) is obtained.

Here, if the probability distributions of the random variable X = { _x1 , _x2 ,..., _xn } and the random variable Y = { _y1 , _y2 ,..., _yn } are _Px (x) and _Py (y), respectively, the mutual information I(X;Y) can be calculated using equation (3).

P _XY (x, y) in equation (3) represents the joint probability, which indicates the relative frequency when the sensor data from sensors 12 and 13 are included in each class at the same time.

For example, let us assume that random variable X = Y and we want to find the joint probability P _XY (x ₁ , y ₁ ) for class x ₁ = y ₁ = [-0.4, -0.38]. In this case, if the two measured values are -0.39 and -0.385 respectively at time t = 0.1 seconds, the frequency is 1, and the value obtained by dividing this by the number of sample data is the joint probability P _XY (x ₁ , y ₁ ).

S16: The monitoring unit monitors the numerical changes in the mutual information I(X;Y) calculated in S15, and diagnoses whether there is an abnormality or a sign of an abnormality each time the mutual information I(X;Y) is calculated.

Here, if the numerical change in the mutual information I(X;Y) exceeds a preset threshold, it is determined that the correlation between the sensor data from sensors 12 and 13 has collapsed, and an abnormality has occurred in one of the sensor data. This threshold is also determined based on an empirically obtained range of change.

For example, according to the output results in Figure 8(c), when the mutual information I(X;Y) is calculated with random variable X = Y, the mutual information I does not change significantly up to 4 seconds, but the value begins to increase after 4 seconds.

In this case, as shown by D5 in Figure 8(c), a change exceeding the threshold value was detected after 5 seconds, and an abnormality was detected in one of the sensor data. Based on this detection, the rotating machine 11 is diagnosed as having an abnormality or a sign of an abnormality, and the diagnosis result is displayed on the display device. Note that the threshold value can also be set in stages for each abnormality/sign of an abnormality, and each abnormality/sign of an abnormality can be displayed on the display device.

S17: The KLD conversion unit evaluates the sensor data from the sensors 12 and 13 using the KLDs of Non-Patent Documents 3 to 5. Specifically, the KLD conversion unit generates a reference probability distribution from past probability distributions previously accumulated in the data measurement device 14, records this, and uses it as input. Here, we will explain the case where the reference probability distribution P _X ^* (x) is generated from the past distribution of the sensor 12, as shown in Figure 9(a).

9(b), a probability distribution P Y (x) is obtained at one-second intervals to be compared with the probability distribution P _X ^* (x). In this case, the probability distribution P _Y (x) uses not only the probability distribution of a different sensor, but also the probability distribution of the same sensor as the probability distribution _{P X} ^* (x).

The KLD is then calculated based on the acquired probability distribution P _Y (x) and the reference probability distribution P _X ^* (x). Figure 9(c) shows the state in which the probability distribution P _Y ⁽ x) of sensor 13 is compared with the probability distribution P _{X *} (x) as a probability distribution of a different sensor.

In this case, P _Y (x) is the same random variable as P _X ^* (x), that is, the probability distribution of sensor 13 consisting of the random variable "X = Y = {x ₁ , x ₂ , ... x _n } = {y ₁ , y ₂ , ... _yn }". Therefore, P _Y (x) is substantially the same as P _Y (y) in S12 and S13, and the probability distribution P _Y (x) of sensor 13 output from the probability distribution conversion unit can be used as is as input. Then, first, "KLD(D _KL (X∥Y))" is found using equation (4).

Next, the KLD has asymmetry and shows a different value. This allows two variables to be obtained from the set of probability distributions P _X ^* (x) and P _Y (x). The JSD of Non-Patent Document 1, which attempts to offset the asymmetry of the KLD, can only obtain one numerical value from the set of probability distributions, and requires relatively less information to make a diagnosis than the KLD. This embodiment utilizes the asymmetry of the KLD for diagnosis, making it possible to make a diagnosis based on a wealth of information. Furthermore, it enables a diagnosis based on a wealth of information compared to the "Wasserstein" distance of Non-Patent Document 2. Here, "KLD(D _KL (Y∥X))", which swaps the probability distributions of Equation (4), is calculated using Equation (5).

9(c), i.e., the two variables "D _KL (X∥Y), D _KL (Y∥X)" are obtained. Similar processing is performed by replacing the probability distribution P _X ⁽ x) of sensor 12 with the probability distribution P _Y (x) to be compared, which is the probability distribution of the same sensor as P _X *(x).

In this embodiment, as an example, the sensor 12 is used as the target and the KLD is calculated based on the probability distribution after one second shown in Figure 9(a) , but the reference probability distribution can also be set as the target for sensor 13. In this case, the reference probability distribution generated from the past probability distribution of sensor 13 is replaced with P _X ^* (x), and the probability distributions of sensors 12 and 13 are replaced with the probability distribution P _Y (x) to be compared.

S18: The monitoring unit monitors each of the two variables calculated in S17 and diagnoses the rotating machine 11 based on deviations of each of the two variables from their normal distribution. The normal distribution is determined in advance from the values of the two variables by executing S11, S12, and S17 while the rotating machine 11 is operating normally.

In Figure 9(c), D6 indicates a normal distribution, and D7 indicates the two variables. In this case, the two variables D7 are detected as outliers from the normal distribution D6 and are therefore recognized as an abnormal state. As a result, the rotating machine 11 is diagnosed as having an abnormality or signs of an abnormality, and the diagnosis result is displayed on the display device. This outlier detection can be performed by checking whether the value deviates from the normal distribution by more than a threshold value, and the threshold value can be set in stages for each abnormality/sign of an abnormality.

This type of equipment diagnosis makes it possible to diagnose the relationship between sensor data collected by sensors 12 and 13 based on changes in entropy caused by the same random variable, and to further evaluate correlation by detecting changes in mutual information and KLD outliers. This makes it easier to perform diagnosis using data from multiple sensors.

The present invention is not limited to the above-described embodiment, and can be modified and implemented within the scope of the claims. For example, the sensors are not limited to the acceleration sensor 12 and current sensor 13, and other sensors (level sensors, flow meters, etc.) may also be used. Furthermore, it is not necessary to perform a unique diagnosis for each sensor using entropy H, and diagnosis without using the asymmetry of KLD is also possible.

9... Equipment diagnostic system 10... Equipment diagnostic device 11... Rotating machine (equipment)
12...Acceleration sensor 13...Current sensor 14...Data measuring instrument

Claims (10)

A device that installs a plurality of sensors in a facility to be diagnosed, and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data,
a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each sensor, divides the set distribution interval into a number of classes that indicate a preset random variable, and converts the set distribution interval into a probability distribution according to the number of measurement values of each class;
a mutual information conversion unit that calculates mutual information for the same time from the probability distribution of each sensor converted by the probability distribution conversion unit using equation (3);

P _X (x): Probability distribution of random variable X = {x ₁ , x ₂ , . . . x _n } of one sensor P _Y (y): Probability distribution of random variable Y = {y ₁ , y ₂ , . . . _yn } of the other sensor where random variable X = random variable Y
I(X;Y): Mutual information of P _X (x) and P _Y (y) P _XY (x, y): Joint probability (relative frequency when data from one sensor and data from another sensor are included in each class at the same time)
Equipped with
an equipment diagnosis device that determines a correlation between data from one of the sensors and data from the other of the sensors based on a change in a value of mutual information calculated by the mutual information transform unit, and diagnoses the equipment in accordance with a result of the determination. A device that installs a plurality of sensors in a facility to be diagnosed, and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data,
a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each sensor, divides the set distribution interval into a number of classes that indicate a preset random variable, and converts the set distribution interval into a probability distribution according to the number of measurement values of each class;
generating a reference probability distribution P _X ^* (x) from the probability distribution;
A probability distribution P _Y (x) to be compared with the same random variable " _X ⁼ Y = {x 1 , x _{2 ,} . . . x _{n }} " as the probability distribution P _X ^* (x) is obtained from the probability distributions P _X (x) and P Y (y) of each sensor obtained at the same intervals as the probability distribution P _{X *} (x);
a KLD conversion unit that converts the probability distributions P _X ^* (x) and P _Y (x) into the KLD "Kullback-Leibler divergence/D _KL (X∥Y)" of equation (4);
Equipped with

Depending on whether the transformed KLD is detected as an outlier away from the normal distribution in the scatter plot,
An equipment diagnosis device for diagnosing the equipment. The KLD conversion unit further exchanges the probability distribution P _X ^* (x) with the P _Y (x) and converts it into the KLD "D _KL (Y∥X)" of equation (5), while

Depending on whether the KLDs of equations (4) and (5) are both detected as outliers away from the normal distribution in the scatter plot,
3. An equipment diagnosis device according to claim 2, wherein said equipment is diagnosed. an entropy conversion unit that converts the probability distribution of each sensor converted by the probability distribution conversion unit into an entropy for each arbitrary time period using equations (1) and (2);


H(X): Entropy of one sensor H(Y): Entropy of the other sensor
Either one of the correlation determination described in claim 1 and the outlier detection described in claim 2;
a threshold determination as to whether the entropy of each sensor has changed by more than a threshold;
3. The equipment diagnosis device according to claim 1, wherein the equipment is diagnosed by: A device that installs a plurality of sensors in a facility to be diagnosed, and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data,
a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each of the sensors, divides the set distribution interval into a number of classes that indicate a preset random variable, and converts the set distribution interval into a probability distribution according to the number of measurement values of each class;
The system comprises a mutual information converter and a KLD converter according to claims 1 and 2,
The correlation determination according to claim 1;
The outlier detection according to claim 2;
An equipment diagnosis device characterized by diagnosing the equipment by A method executed by a device that installs a plurality of sensors in a facility to be diagnosed and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data, comprising:
a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each sensor, dividing the set distribution interval into a number of classes representing a preset random variable, and converting the set distribution interval into a probability distribution according to the number of measurement values of each class;
a mutual information conversion step of calculating mutual information for the same time from the probability distribution of each sensor converted in the probability distribution conversion step using equation (3);

P _X (x): Probability distribution of random variable X = {x ₁ , x ₂ , . . . x _n } of one sensor P _Y (y): Probability distribution of random variable Y = {y ₁ , y ₂ , . . . _yn } of the other sensor where random variable X = random variable Y
I(X;Y): Mutual information of P _X (x) and P _Y (y) P _XY (x, y): Joint probability (relative frequency when data from one sensor and data from another sensor are included in each class at the same time)
a diagnosis step of determining a correlation between data from one of the sensors and data from the other sensor based on a change in the value of the mutual information calculated in the mutual information conversion step, and diagnosing the equipment in accordance with the result of the determination;
An equipment diagnosis method comprising: A method executed by a device that installs a plurality of sensors in a facility to be diagnosed and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data, comprising:
a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each sensor, dividing the set distribution interval into a number of classes representing a preset random variable, and converting the set distribution interval into a probability distribution according to the number of measurement values of each class;
generating a reference probability distribution P _X ^* (x) from the probability distribution;
A probability distribution P _Y (x) to be compared with the same random variable " _X ⁼ Y = {x 1 , x _{2 ,} . . . x _{n }} " as the probability distribution P _X ^* (x) is obtained from the probability distributions P _X (x) and P Y (y) of each sensor obtained at the same intervals as the probability distribution P _{X *} (x);
a KLD conversion step of converting the probability distributions P _X ^* (x) and P _Y (x) into the KLD "Kullback-Leibler divergence/D _KL (X∥Y)" of equation (4);

a diagnosis step of diagnosing the equipment based on whether the KLD converted in the KLD conversion step is detected as an outlier away from a normal distribution in a scatter plot;
An equipment diagnosis method comprising: In the KLD conversion step, the probability distribution P _X ^* (x) and the P _Y (x) are further exchanged and converted into the KLD "D _KL (Y∥X)" of equation (5), while

In the diagnosis step, depending on whether or not both of the KLDs of the formulas (4) and (5) are detected as outliers away from the normal distribution in the scatter plot,
8. The equipment diagnosis method according to claim 7, further comprising diagnosing the equipment. an entropy conversion step of converting the probability distribution of each sensor converted in the probability distribution conversion step into an entropy for each arbitrary time period using equations (1) and (2);
Furthermore,


H(X): Entropy of one sensor H(Y): Entropy of the other sensor The diagnosis step is
Either one of the correlation determination described in claim 6 and the outlier detection described in claim 7;
a threshold determination as to whether the entropy of each sensor has changed by more than a threshold;
8. The equipment diagnosis method according to claim 6, wherein the equipment is diagnosed by: A method executed by a device that installs a plurality of sensors in a facility to be diagnosed and diagnoses the facility using time-series data of an arbitrary time width of each of the sensors as measurement data, comprising:
a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each sensor, dividing the set distribution interval into a number of classes representing a preset random variable, and converting the set distribution interval into a probability distribution according to the number of measurement values of each class;
Execute the mutual information conversion step and the KLD conversion step according to claims 6 and 7,
The correlation determination according to claim 6.
The outlier detection according to claim 7;
An equipment diagnosis method comprising: diagnosing the equipment by