Presentation Outline
What is a “Inferential Sensor”? Background Industrial application Homeland security EDEN Network Water-level Inferential Sensor Challenges	[larger image]

Tough Environment to Monitor
Emissions regulations require measurements of effluent gases Smoke stack burns up probes Need alternative to “hard” sensors	[larger image]

Hard Sensor vs. Inferential Sensor

[larger image]

Virtual sensor replaces actual sensor Temporary gage smoke stack Operate industry to cover range of emissions Develop model of emissions based on operations Model becomes the “Inferential Sensor"

Industrial Application: Production Emission Monitoring System

[larger image]

Production Control System Components. A Data Historian is a special database designed to hold massive amounts of process and laboratory time series data.

Production Emission Model

[larger image]

Predicted Emissions History for a Manufactured Product.

Inferential Sensor Architecture

[larger image]

PEMS Architecture

Industrial Benefits

Optimizes Manufacturing Processes Documents Continuous Compliance Lower Cost - In some situations, a proactive intelligent system can partially or fully replace passive back end controls to reduce capital and long-term operating expenses Most Advantageous Permitting - because it actively prevents pollution Works with Existing or New Production Controls

If it is good enough for Industry...

Use similar approach for real-time data Develop models to predict real-time data Use predictions as "inferential-sensor" to: QA/QC hard sensor Provide accurate estimates for hard sensor Provide redundant signal

Everglades

[larger image]

Water-depth and Ecosystems

[larger image]

Differences of 1 ft can change vegetation communities

Everglades Depth Estimation Network (EDEN)
Grid up the Everglades (400 m) Measure elevation of each grid Create digital elevation model (DEM) Create surface-water map Water depths = Surface-Water Map - DEM Goal: generate real-time water-surface and water-depth maps	[larger image]

EDEN - Model Integration

[larger image]

EDEN Gaging Network
Agencies: USGS S. Florida Water Management District Big Cypress National Preserve Everglades National Park	[larger image]

EDEN Water-Surface Maps
Daily medians computed from hourly data Water surface maps generated using radial basis function (RBF) in GIS	[larger image]

EDEN Water-Surface Map
Bad values creates erroneous maps -	[larger image]

Problem Need to minimize missing and erroneous data Approach Develop "inferential" sensors for redundant signal

[larger image]

Hypothetical Case
[larger image]	Create model (Inferential Sensor) for Site B using Site A as an input Decide when to use Inferential Sensor instead of gage data Actual application would be for 253 stations

Hypothetical Case: Gage Data

[larger image]

Hypothetical Case: Inferential Sensor

[larger image]

Hypothetical Case: When to use Inferential Sensor?

[larger image]

Hypothetical Case: When to use Inferential Sensor?

[larger image]

Hypothetical Case - comments

Issue of model accuracy Immediate benefit for missing data Made the assumption that Site A was correct Example used daily data Use inferential sensor on hourly data Compare daily medians (used for EDEN maps) What if data for Site A is missing? Issues are magnified when dealing with a network of 253 gages

Initial Approach: Data Evaluation

Need to know what data is good Set of filters to evaluate data quality Robust series of thresholds Differences with other gages Time delays and moving window averages Time derivatives Rate of change over various time intervals Create subset of good data

Initial Approach: Model Development

One Approach - Canned Models Create multiple models for a gage Set priority for model to use depending on available data Large number of models Not all combinations of gages would be addressed

Initial Approach: Model Development

Second Approach - Model on the Fly Subset of good data, determine input data with the highest correlation Develop models based on available data Issue of evaluation of models More complex programming than first approach

EDEN Inferential Sensor Architecture
Data from NWIS - 253 sites
Data Evaluation - robust filters
Inferential Sensor Models	Output Log File
Data Fill and Replacement
Data to EDEN server

Summary
Inferential Sensor may provide an approach for: Real-time QA/QC Redundant signal Challenges Identifying good data Highly accurate models Managing number of models Develop prototype for EDEN Stay tuned	[larger image]

Questions?

[larger image]