Presentation is loading. Please wait.

Presentation is loading. Please wait.

Jorge Ortiz.  Metadata verification  Scalable anomaly detection.

Similar presentations


Presentation on theme: "Jorge Ortiz.  Metadata verification  Scalable anomaly detection."— Presentation transcript:

1 Jorge Ortiz

2  Metadata verification  Scalable anomaly detection

3 Chiller Pump Chiller Pump AHU SFEF Vent Zone

4 Chiller Pump Chiller Pump AHU SFEF Vent Zone System Space

5  Geometric  Placement, associations  Functional  Temperature, pressure, flow, etc.  Semantic  Electrical device taxonomy  Ownership

6 Current Our work

7  Are the geometric (spatial) associations correct?  Are all the sensors with the same spatial grouping in the same location?  Sensors can be moved or replaced  Contractor mislabels point in BMS  How can the sensor data guide this process? SODA4 R520 __ART

8  Sensor streams driven by same phenomena  Common trend ineffective at uncovering relationships

9  Each row/column is a location in the building  Each location has one or more sensors  Cell (i,j) is the average device pairwise correlation between sensors at locations i and j

10  Approach used for finding underlying data trends  Algorithm for decomposing signals in the time domain of non-stationary, non-linear signals  Similar to FFT, PCA but yields characteristic time and frequency scales  Output “Intrinsic mode functions”  Combination of underlying signal in the same time scale

11

12

13 Compare the EHP to 674 other sensors: EMD helps us to discriminate un/related sensors **Suggests Geometric Verification is possible**

14  Mislabeled “type” information of a data stream  Fault detection  Strip, bind, and search process

15  Difficult for building managers to know where to start to look for problems  Which devices? Locations? Patterns? Time interval?  Key Observation  Devices are used simultaneous in the same way  Typically usage times/patterns are tightly un/coupled ▪ Example: ▪ Lights and HVAC during the day  Basic assumption  Normal usage is efficient.  Pairwise correlation analysis of sensor traces  Uncover usage relationships between devices

16

17  Construct reference matrix for each time-reference interval  For new data points, compute l  Identifying outliers  Median Absolute Deviation p=4, b=1.4826

18  High power usage  Alarms corresponding to electricity waste  Lower power usage  Alarms representing abnormal low electricity consumption  Punctual  Short increase/decrease in electricity consumption  Missing data  Possible sensor failure  Other  unknown

19 AC On All Night Lights On All Night AC Not On During Office Hours

20 Possible Chiller dysfunction Change in power usage pattern Simultaneous heating and cooling Normal 18 days, 2500 kWh

21  2 research papers in collaboration with U. of Tokyo  Internet of Thing Workshop @IPSN 2012  IPSN 2013 (April)  Web tool that finds anomalies from data uploads  Upcoming release

22  Value verification  Model-based verification, model validation  Standard representation with embedded confidence parameters for MPC


Download ppt "Jorge Ortiz.  Metadata verification  Scalable anomaly detection."

Similar presentations


Ads by Google