ANOMALY DETECTING APPARATUS

• US 20130024172A1
• Filed: 09/27/2012
• Published: 01/24/2013
• Est. Priority Date: 03/30/2010
• Status: Active Grant

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: SUYAMA, Akihiro, HANADA, Yuuichi
• Granted Patent: US 8,577,649 B2
• Time in Patent Office: Days
• US Class Current: 703/2
• CPC Class Codes: G05B 2219/24042 Signature analysis, compare recorded with current data, if error then alarm
• CPC Class Codes: G05B 23/0232 based on qualitative trend analysis, e.g. system evolution
• CPC Class Codes: G05B 23/0254 based on a quantitative model, e.g. mathematical relationships between inputs and outputs; functions: observer, Kalman filter, residual calculation, Neural Networks

Abstract

There is provided an apparatus, including a first storage storing first time-series data including instruction values given to first to K-th control devices and measured values from first to K-th sensors during a first period, a model optimizer, for each of combinations of two of the control devices, generating a diagnostic model instance of a predetermined target model and obtaining an optimized diagnostic model instance in which parameters of the diagnostic mode instance are identified, a second storage storing second time-series data acquired during a second period, a calculator, for each of the optimized diagnostic model instances, calculating a determination score for each of the control devices and sensors using the first and second time-series data, and a determiner determining presence or absence of an abnormality for each of the control devices and the sensors based on each determination score.

Claims

1. An anomaly detecting apparatus for first to K-th control devices and first to K-th sensors, the control devices being arranged at different locations on a medium path that control output of medium flowing in the medium path according to respective externally given instruction values and the sensors measuring output volumes of the medium controlled by the control devices, comprising:

a target model storage storing a target model defined by (Ci+β

i(Xi))×

(Qi/Qj)̂

2−

(Cj+β

j(Xj), where “

Xi”

is a first observed variable, “

β

i(Xi)”

is a first unknown characteristic term that is an unknown function having input variable as “

Xi,”

“

Ci”

is a first constant term, “

Qi”

is a second observed variable, “

Xj”

is a third observed variable, “

β

j(Xj)”

is a second unknown characteristic term that is an unknown function having input variable as “

Xj,”

“

Cj”

is a second constant term, and “

Qj”

is a fourth observed variable;

a first storage storing first time-series data each including the instruction values given to the first to K-th control devices and measured values from the first to K-th sensors at predetermined time intervals during a first period;

a model optimizer configured to generate, for each of combinations of two of the first to K-th control devices, a diagnostic model instance by assigning the instruction value for one of the two control devices to the “

Xi”

of the target model, an unknown constant corresponding to the location of the one control device to the “

Ci,”

the measured value from the sensor corresponding to the one control device to the “

Qi,”

the instruction value for the other control device to the “

Xj,”

an unknown constant corresponding to the location of the other control device to the “

Cj,” and

the measured value from the sensor corresponding to the other control device to the “

Qj” and

obtain each optimized diagnostic model instance in which the “

β

i(Xi),”

the “

Ci,”

the “

β

j(Xj),” and

the “

Cj”

are identified using the first time-series data so as to minimize the squares of values of all of each diagnostic model instance, where the “

β

i(Xi)”

is a function that takes values for sections partitioned in a domain of the “

Xi” and

the “

β

j(Xj)”

is a function that takes values for sections partitioned in a domain of the “

Xj”

;

a second storage storing second time-series data each including the instruction values given to the first to K-th control devices and measured values from the first to K-th sensors at predetermined time intervals during a second period different from the first period;

a third storage storing accuracy information, representing maximum observation errors of the first to K-th sensors and maximum response errors of the first to K-th control devices for the instruction values;

a determination score calculator configured to, for each optimized diagnostic model instance,

obtain first residuals being values calculated from the optimized diagnostic model instance based on each piece of the first time-series data and generate a first probability distribution which is a probability distribution of the first residuals,

obtain second residuals values calculated from the optimized diagnostic model instance based on each piece of the second time-series data and generate a second probability distribution which is a probability distribution of the second residuals,

change, for each of the two control devices and two sensors based on which the optimized diagnostic model instance is generated, each corresponding value in each piece of the first time-series data according to the accuracy information, obtain third residuals being values calculated from the optimized diagnostic model instance based on each corresponding changed piece of time-series data, and generate each corresponding third probability distribution which is a probability distribution of the third residuals, and

generate a first extended KL (Kulback-Leibler) divergence representing a distance between the second and first probability distributions and second extended KL divergences each representing a distance between the second probability distribution and each third probability distribution and calculate a determination score according to the ratio of the first extended KL divergence to the sum of the first extended KL divergence and each one of the second extended KL divergences, respectively; and

a comprehensive abnormality determiner configured to obtain, for each of the first to K-th control devices and the first to K-th sensors, a comprehensive score which is a minimum value of determination scores obtained for each of the first to K-th control devices and the first to K-th sensors and, if the comprehensive score of any of the control devices and sensors is larger than a first threshold value, determine that an abnormality has occurred in the control device or the sensor.

2. The apparatus according to claim 1, wherein

the model optimizer counts, for each of the first to K-th control devices, a number of pieces of data for each of the sections, the instruction value of which belongs to each of the sections, and sets a confidence value that increases with an increase in the number of pieces of data for each of the sections, and

the determination score calculator

calculates reliability by multiplying confidence values of two sections which cover values of the “

Xi” and

the “

Xj”

in each piece of first time-series data and generates the first probability distribution using only the first residuals that have been obtained from ones of the pieces of data for each of which the reliability not less than a second threshold value has been calculated,

calculates reliability by multiplying confidence values of two sections which cover values of the “

Xi” and

the “

Xj”

in each piece of second time-series data and generates the second probability distribution using only the second residuals that have been obtained from ones of the pieces of data for each of which the reliability not less than the second threshold value has been calculated, and

calculates reliability by multiplying confidence values of two sections which cover values of the “

Xi” and

the “

Xj”

in each piece of changed time-series data and generates the third probability distribution using only the third residuals that have been obtained from ones of the pieces of data for each of which the reliability not less than the second threshold value is calculated.

3. The apparatus according to claim 1, wherein

the model optimizer further generates a diagnostic model instance by assigning the instruction value for the other control device to the “

Xi,”

the unknown constant related to the other control device to the “

Ci,”

the measured value from the sensor corresponding to the other control device to the “

Qi,”

the instruction value for the one control device to the “

Xj,”

the unknown constant related to the one control device to the “

Cj,” and

the measured value from the sensor corresponding to the one control device to the “

Qj.”

4. The apparatus according to claim 1, wherein

the determination score calculator

calculates squares of the first residuals and generates a probability distribution of the squares of the first residuals as the first probability distribution,

calculates squares of the second residuals and generates a probability distribution of the squares of the second residuals as the second probability distribution, and

calculates squares of the third residuals and generates a probability distribution of the squares of the third residuals as the third probability distribution.

5. The apparatus according to claim 1, wherein

the determination score calculator calculates the determination score according to the ratio of the second extended KL divergence to the sum of the first extended KL divergence and each one of the second extended KL divergences, respectively, and

the comprehensive abnormality determiner obtains a maximum value of each determination score as the comprehensive score and, if the comprehensive score of any of the control devices and sensors is less than the first threshold value, determines that an abnormality has occurred in the control device or sensor.

6. The apparatus according to claim 1, wherein

the target model storage stores, as the target model, (Ci+β

i(Xi))×

f(Qi, Qj)−

(Cj+β

j(Xj)) instead of the (Ci+β

i(Xi))×

(Qi/Qj)̂

2−

(Cj+β

j(Xj), where (f(Qi, Qj) is a predetermined function whose input variables are “

Qi” and

“

Qj”

.

7. The apparatus according to claim 1, wherein the first to K-th control devices are each a variable air volume terminal or a constant air volume terminal, the instruction values given to the first to K-th control devices are each a damper position for a valve of the variable air volume terminal or the constant air volume terminal, and the first to K-th sensors are flow sensors.

8. The apparatus according to claim 1, wherein the first to K-th control devices are water regulating devices, the instruction values given to the first to K-th control devices are each a damper position for a valve of the water regulating device, and the first to K-th sensors are water volume sensors.