Error detection method and its system for early detection of errors in a planar or facilities

• US 8,630,962 B2
• Filed: 05/29/2009
• Issued: 01/14/2014
• Est. Priority Date: 10/09/2008
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Maeda, Shunji, Shibuya, Hisae
• Primary Examiner(s): CHANG, LI WU
• Application Number:
  US13/057,831
• Publication Number:
  US 20110191076A1
• Time in Patent Office: 1,691 Days
• Field of Search: None
• US Class Current: 706/12
• 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
• CPC Class Codes: G06F 18/21375 involving differential geometry, e.g. embedding of pattern manifold
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
• CPC Class Codes: G06F 2218/12 Classification; Matching

Abstract

Provided are a method which permits complete training data and data with added errors, and enables the early and accurate discovery of errors in facilities such as a plant, and a system thereof. To achieve the objectives, (1) the behavior of temporal data is observed over time, and the trace is divided into clusters; (2) the divided cluster groups are modeled in sub spaces, and the discrepancy values are calculated as errors candidates; (3) the training data are used (compare, reference, etc.) for reference to determine the state transitions caused by the changes over time, the environmental changes, the maintenance (parts replacement), and the operation states; and (4) the modeling is a sub space method such as regression analysis or projection distance method of every N data removing N data items, (N=0, 1, 2, . . . ) (for example, when N=1, one error data item is considered to have been added, this data is removed, then the modeling is performed), or a local sub space method. Linear fitting in regression analysis is equivalent to the lowest order regression analysis.

Claims

1. An error detection method for early detection of errors in a plant or facilities, comprising of steps of:

acquiring data from a plurality of sensors;

dividing a trace in a data space into a plurality of clusters over time on the basis of a temporal change of the data;

wherein a trace in a data space represents the sensor data acquired over a period of time;

wherein a data space represents attributes of sensor data;

wherein each cluster corresponds to an operation state of the plant or facilities;

for each operation state, modeling the corresponding divided clusters by a sub space method; and

calculating a discrepancy value based on the models of derived clusters as an error candidate.

2. The error detection method according to claim 1, wherein in the modeling, training data including past data from the plurality of sensors is modeled to create a reference model, and the reference model is parallel-used to calculate an discrepancy value as an error candidate.

3. The error detection method according to claim 1, wherein the modeling is performed by use of either regression analysis of every N data removing N data items (N=0, 1, 2, . . . ) or the sub space method.

4. The error detection method according to claim 3, wherein the sub space method is a projection distance method, a CLAFIC method or a local sub space method for measuring an area around observation data.

5. The error detection method according to claim 4, wherein the detection of the discrepancy value is performed by integration of outputs in the regression analysis, the projection distance method, the CLAFIC method or the local sub space method for measuring an area around observation data.

6. An error detection method for early detection of errors in a plant or facilities, comprising of steps of:

acquiring data from a plurality of sensors;

dividing a trace in a data space into a plurality of clusters over time on the basis of a temporal change of the data;

wherein a trace in a data space represents the sensor data acquired over a period of time;

wherein a data space represents attributes of sensor data;

wherein each cluster corresponds to an operation state of the plant or facilities;

for each operation state, modeling the corresponding divided clusters by a sub space method; and

displaying the data on a time-series signal to make the divided clusters evident to visualize a state.

7. The error detection method according to claim 6, wherein a first report is output by error detection performed on the clusters on the basis of past data from the plurality of sensors, and a second report is output by error detection performed on the clusters on the basis of current data of the plurality of sensors.

8. An error detection system for early detection of errors in a plant or facilities, comprising:

a hardware processor;

a data acquiring unit that acquires data from a plurality of sensors;

a trace division unit that divides a trace in a data space into a plurality of clusters over time on the basis of a temporal change of the data acquired by the data acquiring unit;

wherein a trace in a data space represents the sensor data acquired over a period of time;

wherein a data space represents attributes of sensor data;

wherein each cluster corresponds to an operation state of the plant or facilities;

a modeling unit that performs modeling in a sub space on the divided clusters corresponding to each operation state which is divided by the trace division unit; and

an error detection unit that calculates an discrepancy value as an error candidate based on the models of derived clusters which is modeled by the modeling unit.

9. The error detection system according to claim 8, comprising a training database unit that accumulates past data from the plurality of sensors by the data acquiring unit, and models the past data to create a reference model,

wherein the reference model in the training database unit is parallel-used to calculate an discrepancy value as an error candidate.

10. The error detection system according to claim 8, wherein the modeling is performed by use of either regression analysis of every N data removing N data items (N=0, 1, 2, . . . ) or a sub space method.

11. The error detection system according to claim 8, wherein the sub space method is a projection distance method, a CLAFIS method or a local sub space method for measuring an area around observation data.

12. The error detection system according to claim 8, wherein the detection of the discrepancy value is performed by integration of outputs in the regression analysis, the projection distance method, the CLAFIS method or the local sub space method for measuring an area around observation data.

13. An error detection system for early detection of errors in a plant or facilities, comprising:

a hardware processor;

a data acquiring unit that acquires data from a plurality of sensors;

a trace division unit that divides a trace in a data space into a plurality of clusters over time on the basis of a temporal change of the data acquired by the data acquiring unit;

wherein a trace in a data space represents the sensor data acquired over a period of time;

wherein a data space represents attributes of sensor data;

wherein each cluster corresponds to an operation state of the plant or facilities;

a modeling unit that performs modeling in a sub space on the divided clusters corresponding to each operation state which is divided by the trace division unit;

an error detection unit that calculates an discrepancy value as an error candidate based on the models of derived clusters which is modeled by the modeling unit; and

a data display unit which displays the data on a time-series signal to make the divided clusters which is divided by the trace division unit evident to visualize a state.

14. The error detection system according to claim 13, wherein a first report is output by error detection performed on the clusters on the basis of the past data from the plurality of sensors, and a second report is output by error detection performed on the clusters on the basis of current data of the plurality of sensors.