K. Kolomvatsos1, C. Anagnostopoulos2, and

S. Hadjiefthymiades1

An Efficient Environmental Monitoring System adopting Data Fusion,

Prediction & Fuzzy Logic

1University of Athens, Department of Informatics & Telecommunications

2Department of Informatics, Ionian University

6th International Conference on Information, Intelligence, Systems and Applications July 2015

Corfu, Greece

OutlineEnvironmental MonitoringThe Proposed MechanismThe Fusion ComponentThe Prediction ComponentThe Fuzzy Logic ControllerExperimental Results

Environmental MonitoringChanges in the environment should be

immediately identified and specific solutions should be applied.

Need: pro-active methods that immediately respond to any change in the environment’s characteristics.

A number of sensors could monitor specific areas and a Central System (CS) could reason over the observed values and support decision making.

Many research and commercial efforts adopt: sensors observing a specific phenomenon (e.g.,

temperature, humidity, water level), and,intelligent systems that respond to the identification

of events (e.g., fire, flood)

The Proposed Mechanism (1/3)We propose a mechanism that combines data

fusion techniques, prediction (regression) and Fuzzy Logic (FL).

The proposed mechanism is a decision making tool for the identification of

environmental events.builds on top of measurements performed by a

number of sensors to provide immediate responses to any observed abnormality.

derives knowledge from the group of sensors.aggregates the reported measurements and

reasons over the opinion of the group (on the occurrence of the specific event).

The Proposed Mechanism (2/3)We consider a group of N sensors.Sensors observe the same phenomenon and their

reports are TXed at pre-defined intervals to the CS.When the CS receives sensor readings, it decides if

an event has occurred and, then, triggers the respective alert.

The CS is not based only on a single sensor observations, as false alarms could undermine robustness.

Sensors could report invalid observations due to various reasons (intrinsic, extrinsic) .

UsersCS

Alerts - Data

Sensor 1

Sensor 2

Sensor 3

Sensor N

...

The Proposed Mechanism (3/3) In the proposed CS, a number of components

are responsible to manage the incoming data and derive the appropriate decision.

The adopted components are:a Fusion Component (FC) that undertakes the

responsibility of eliminating the outlier data and providing the final fused measurement,

a Prediction Component (PC) that, based on the fused measurement historical values, derives the predicted future fused measurement and,

an FL Controller (FLC) that gets the results of FC and PC components and derives the Degree of Danger (DoD). The DoD provides a view on the existing danger based

on the current measurements and the predicted value. Sensor 1

Sensor 2

Sensor 3

Sensor N

...

FC

PC

FLC

CS

H

DoD

The Fusion Component (1/4)Data fusion combines contextual data from all sensors to

derive reliable fused measurements.We adopt the cumulative sum (CumSum) algorithm for

outliers detection over all measurements and the linear opinion pool algorithm for deriving the final aggregated value.

CumSum detects if there is any change in the distribution of a contextual time series corresponding to a sensor.

The algorithm is a change-point detection technique based on the cumulative sum of the differences between the current value at instance t and the overall average up to t.

We adopt a 2-side detection scheme where an observation is inferred as an outlier when it lies above a target threshold h+ or below a target threshold h-

The Fusion Component (2/4)Two signals are the output of the CumSum algorithm:

the above detection signal and the below detection signal.

When the time series deviates from the thresholds, detection signals are set to 1.

The detected outliers are eliminated and, thus, the mechanism (at instance t) is based on the opinion (measurements) of those sensors that do not produce outliers.

The proposed MS adopts a linear opinion pool scheme for the remaining values.

The Fusion Component (3/4)The linear opinion pool is a standard approach to

combine experts’ opinion through a weighted linear average.

The final aim is to combine singe experts’ opinion to produce the opinion of the group.

We apply specific weights in each expert to pay more / less attention on each opinion affecting more / less the final result.

The fused measurement f = F(x1, x2, …, xN) is the opinion pool based on the pooling operator F over the measurements (opinions).

We adopt a weighted linear average i.e.,

, where wi is the weight associated with the

opinion of sensor Si which does not produce an

outlier. m is the number of sensors derived by the

CumSum algorithm..

Nm

1iiixwf

The Fusion Component (4/4)We define Ci as the confidence that the CS

has on sensor Si in successfully fulfilling the assigned task.

Ci could be affected by the sensor state or by historical data.

We calculate each weight as follows:

The CS assigns more weight on the opinion of a sensor having a high confidence value.

Nm

1jj

ii

C

Cw

The Prediction ComponentFor each sensor Si a time ordered set of

past values is maintained.History of the latest observations.We predict the next measurement through

a linear combination of the historical measurements with real-valued prediction coefficients.

The set of coefficients are estimated to minimize the prediction error between the predicted and the actual measurement.

We adopt the Levinson-Durbin* algorithm for coefficient estimation.* A)Durbin J., ‘The fitting of time series models’, Rev. Inst. Int. Stat., v. 28, 1960, pp. 233-243 B) Levinson N., ‘The Wiener RMS error criterion in filter design and prediction’, Journal Math. Phys., v. 25, 1947, pp. 261-278

The FL Controller We developed an FL Controller (FLC), which

is responsible for defining the CS’s reaction to the incoming data

The FL Controller has two inputs:the fused measurement based on current

sensors observations (fv)the predicted measurement based on

historical data (pv) The output is the DoD value. We consider that DoD 1 indicates that the

danger is at high levels; the opposite stands when DoD 0.

We consider three linguistic values: Low, Medium, High.

A Low value represents that the variable takes values close to the lower limit while a High value depicts the case where the variable takes values close to the upper level. A Medium value depicts the case where the variable takes values close to the average.

We consider triangular membership functions.

Experimental Evaluation (1/3)Performance Metrics

Rate of False Alerts (RFA). It represents the rate of false alerts.As RFA → 1, the system results a lot of false alerts. As RFA

→ 0, the system minimizes the rate of false alerts.Index of Alert (IA).

It represents the index of the sensor report that triggers an alert.

DatasetsIntel Berkeley Research Lab dataset – Temperature

values (Intel Lab Data, http://db.csail.mit.edu/labdata/labdata.html)We inject faulty measurements to derive the RFA - P faulty

measurements (P ∈ {1%, 5%, 10%, 20%, 40%}) of the total (15,000 measurements)

Real past flood event as reported by a number of water level sensors (http:// www.pegelonline.wsv.de)

Experimental Evaluation (2/3)We compare the proposed FL system (FLS)

with the following models:The Single Sensor Alerting (SSA) mechanism

It delivers an alert when at least one sensor reports a value over a pre-defined threshold

The Average Measurements Alerting (AMA) mechanismIt produces alerts when the average value of the

sensor reports is over the pre-defined threshold

Experimental Evaluation (3/3)Experimental results

for various P

Experimental results for various N (number of sensors)

Experimental results for IA (5 sensors)

Experimental results for IA (10 sensors)

Model IA

Real Event

45

FLC 44

AMA 88

Model IA

Real Event

45

FLC 40

AMA 45

SSA 34, 35, 36, 39

Thank You!