20
Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of Technology Management Jacobo Sack IPN Representation in Israel Adin Holdings Ltd. Israel Sept. 2014

Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Embed Size (px)

Citation preview

Page 1: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

IPN-ISRAEL WATER WEEK (I2W2)

Eyal Brill Holon institute of Technology, Faculty of Technology Management

Jacobo SackIPN Representation in Israel

Adin Holdings Ltd. IsraelSept. 2014

Page 2: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Real-time measuring of water quality is an essential necessity for the water quality manager.

Few specific on line sensors developed such as:chlorine residualTurbiditypHTemperaturedissolved oxygenelectrical conductivityORP and few others

Page 3: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Detection of water contamination (henceforth DWC) in drinking water systems has received increasing attention over the past years that has brought to the development of:

On line instrumentation real time measurements of water quality parameters Models with ability to detect water quality events (EDS: Event detection systems).

Page 4: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Figure 1a: Typical weekly measurements for Conductivity and Free chlorine.

Page 5: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Figure 2: Event example - Change in

Figure 2 Event example. Change in Conductivity and pH

Page 6: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

The EPA has devoted a program called "EDS Challenge" to advance the development of EDS models

has published some guidelines and conclusions about how the target of detecting contamination may be achieved.

Page 7: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Identifying the occurrence of potential events in water systems, as well as in industrial systems, can be accomplished by detecting rare events.

One of the most popular alternatives for implementing DWC is by using unsupervised machine learning algorithms.

This methodology is similar to detecting credit card fraud transactions.

In the first stage, the learning algorithm uses a "clear data" set to learn about the normal behavior of the system. (A clear data set refers to a data set where no alarms have been generated.)

Page 8: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Once the knowledge about normal behavior has been established, the model has to be tested.

A new data set with "cooked" events is introduced to the algorithm. The algorithm is then asked to classify each of the new records.

Since the tester knows the true answers for all of the records, four possible situations may occur for each answer given by the algorithm:

Page 9: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

True Positive – The algorithm alerted correctly.False Positive – The algorithm alerted mistakenly. True Negative – The algorithm did not alert correctly.False Negative – The algorithm did not alert mistakenly.

Page 10: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

cell definition A combination of multi-dimension intervals and discrete values of the water quality variables.

A record is assigned to a cell if the values of all its continuous variables are included within the boundaries of a cell, and the values of all the discrete variables are equal to the specific values of the cell.

The density of a cell is the number of records assigned to each cell.

Page 11: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

From Dataset to Detection Model

Training the model means running historical data and assigning each record to a single cell.

As a result, some cells will be more populated, while others will be less populated.

More populated cells are depicted as common situations, and thus, are safer.

Cells with less or no records are suspect.

Page 12: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

When an abnormal (rare) situation occurs (for example, the intrusion of contamination into the

water), the cell representing this situation is "painted" as a “bad” cell.

If, in the future, a combination similar to that cell occurs, an alert will be generated.

On the other hand, when a "new" situation occurs, the combination of the parameters can't be assigned to

any cell. In this case, human classification (Domain Expert) is needed.

Page 13: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

The domain expert does the following:

Correlate between measurements and events.

Classify unknown situation into Good or Hazard.

Evaluate models performance and initiate relearning when needed.

Page 14: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Test data set propertiesParameter

Name

Average Standard

Deviation

Intervals Statistical

Low

Statistical

High

Conductivity (ms) 102.3 52.1 50 60.5 251.6

Free chlorine (ppm) 2 0.3 0.1 1.5 2.4

pH 8.9 1.1 0.5 9.2 9.8

Pressure (PSI) 34.9 4.3 5.0 32.0 38.0

TOC 1.2 0.7 0.25 0.5 2.1

Turbidity 0.3 0.8 0.25 0.4 2.1

Temperature 13.3 2.7 5 12.9 23.5

Table 1: Dataset properties

Page 15: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Results – Common combination

Combination Count Density

COND CL pH Press Temp TOC Turbidity

45 2.00 9.3 35 15 0.8 0.49 3,839 1

159 2.11 8.9 35 11 1.7 0.07 3,336 0.869

155 2.14 8.7 35 16 1.6 0.05 3,100 0.808

62 2.01 9.1 35 14 1.0 0.27 2,777 0.723

47 1.99 9.3 35 16 0.9 0.45 2,572 0.670

Page 16: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

FP-FN-TP-TNL – Density

(Probability)

T – Time delay

(minutes)

Total

Alerts

TP FP FN DR

(%)

FR

(%)

1 0.001 15 72 19 52 1 95 72

2 0.005 15 289 20 269 0 100 93

3 0.01 15 333 20 303 0 100 90

4 0.001 30 40 19 20 1 95 50

5 0.005 30 169 19 149 1 95 88

6 0.01 30 196 19 176 1 95 89

7 0.001 45 3 3 25 17 15 0

8 0.005 45 86 7 66 13 35 76

9 0.01 45 110 12 90 8 60 81

10 0.001 34 38 15 18 5 75 47

Page 17: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Concluding Remarks

The current paper presents a demonstrated methodology for the detection of water contamination

events, based on a density table .

A density table is constructed directly from the frequency of data combination occurrences.

Testing the accuracy of the model is achieved by using simulating scenarios

Page 18: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

The methodology presented, produces a model, which enables the detection of most events,

with a reasonable level of false alarms.

The data base analysis made in the example that was used in our case, with data recorded over a period of six months, yielded an alarm every 4.7 days, which is on average, one alarm per week.

This level of false alarms is considered reasonable in a water supply network.

Page 19: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

The main strengths of the methodology presented are:

The ability to detect water quality events without prior engineering knowledge

The method’s simplicity

A relatively good detection ratio with a low rate of false alarms

Page 20: Water Contamination Detection – Methodology and Empirical Results IPN-ISRAEL WATER WEEK (I 2 W 2 ) Eyal Brill Holon institute of Technology, Faculty of

Water Contamination Detection –Methodology and Empirical Results

Thank you