Tetra Tech, Inc. · NASDAQ:TTEK
Tetra Tech — Technical Presentation (December 2011)
17 slides from Tetra Tech's technical presentation, Dec 7, 2011
Slide 1: Rare Event Detection Algorithm - Applicability To Water Quality Monitoring
Rare Event Detection Algorithm - Applicability To Water Quality Monitoring Michael J. Ungs 1
Slide 2: Objectives And Technical Approach
Objectives And Technical Approach ■ SCADA systems record large amounts of water quality data and report anomalies to system operators ■ The problem is to identify significant events requiring operator actions ■ Review existing test methods ■ Explore alternative approaches to enhance on-line event de
Slide 3: Standard Algorithms
Standard Algorithms Linear Prediction Filter Filter and then perform linear trend analysis using previous window of data to predict next measurement. Difference between prediction and measurement is compared to a threshold value. Multivariate Nearest Neighbor Map previous window of data into m-dimen
Slide 4: Nature Of The Recorded Data And Limitations Of Standard Algorithms
Nature Of The Recorded Data And Limitations Of Standard Algorithms Observed Data Standard Assumption Multiple magnitudes of Bell shaped distribution of scale values (Gaussian) Multiple frequencies Linear variation over time superimposed at a fixed frequency Random initiation of scale Patterns assume
Slide 5: Standard Assumption Is A Gaussian Distribution
Standard Assumption Is A Gaussian Distribution Actual Data Contains Multiple Magnitudes Of Scale For Example- A Bivariate Distribution Of Total Organic Carbon 5
Slide 6: Standard Assumption Is A Linear Variation Over Time At A Fixed Frequency
Standard Assumption Is A Linear Variation Over Time At A Fixed Frequency Actual Data Contains Multiple Scales & Harmonics In A Data Set 6
Slide 7: Additional Attributes Observed
Additional Attributes Observed ■ Bivariate and trivariate frequencies ■ Many combinations of values never occur ■ Correlation matrix is sparsely populated, as seen in this NxM matrix 7
Slide 8: Summary Of Data Characteristics And Prior Assumptions
Summary Of Data Characteristics And Prior Assumptions ■ Plot of parameter value as function of time shows multiple harmonics in the data ■ Plot of the probability density function (PDF) of data shows multiple modes in data ■ Basic assumption of central tendency is invalid ■ Many combinations of data
Slide 9: Desired Features of New Algorithm
Desired Features of New Algorithm Ability to detect covariate relationships between two or more parameters, as shown in the EPA contaminate-spike formulas Simultaneously control the number of both false positives and false negatives Need to custom fit algorithm to each site using an automatic fittin
Slide 10: Algorithm To Create Correlation Matrix For Two Parameters
Algorithm To Create Correlation Matrix For Two Parameters Pick historical record of approximately 10,000 values Find minimum and maximum value for each parameter Subdivide ranges of two parameters into N and M intervals, creating a NxM matrix Place each data pair into the matrix, counting the number
Slide 11: False Positive Count
False Positive Count Take an additional 10,000 data values Find the cell of the p-dimensional correlation matrix that corresponds to the new data values Count as a false positive value if there are J or less historical values in that cell: increment as FalsePos +1 Repeat for all 10,000 data pairs 11
Slide 12: False Negative Count
False Negative Count Take the same 10,000 data values Add a contaminate spike to data values using Eddies Contaminate Library with a specified level of dose Compare spiked values to correlation matrix. Count as a false negative if there are K or more historical values in that cell: increment as Fals
Slide 13: Testing Different Combinations
Testing Different Combinations Repeat the above process for a different number of cells for each parameter range, varying cell count between 4 and 15 Repeat all of the above processes for a different combination of two or more water quality parameters The Rare Event Detection Algorithm uses that com
Slide 14: Using Algorithm
Using Algorithm The NxM matrix and paired parameters with the smallest FalsePos+FalseNeg count is periodically re-calculated off-line (tuned every 3-5 months) Each new data value is compared to the NxM matrix. An event flag is triggered if any new data pair falls into a cell with less than J histori
Slide 15: 2-D Analysis Of Rare Event Detection Routine With First Data Set
2-D Analysis Of Rare Event Detection Routine With First Data Set Spike with Best Parameter Number of Number of Dose = 2 Pair False Positives False Negatives A TOC & CL2 16 16 B Temp & Cond 712 7187 C Temp & TOC 626 5415 D Temp & pH 715 6879 E TOC & CL2 15 4 F pH & CL2 9 74 Temp & TOC G 626 5414 H Co
Slide 16: 2-D Analysis Of Rare Event Detection Routine With Second Data Set
2-D Analysis Of Rare Event Detection Routine With Second Data Set Spike with Best Parameter Number of Number of Dose = 2 Pair False Positives False Negatives A TOC & CL2 5 1 B Temp & Cond 4583 4585 C TOC & ORP 0 0 D Cond & pH 3381 3287 E TOC & ORP 0 0 F TOC & ORP 0 0 G TOC & ORP 7 25 H TOC & ORP 0 0
Slide 17: Summary
Summary Rare Event Detection Algortihm works extremely well when: All water quality parameters are actually measured The magnitude of the contaminate spike is sufficiently different from background Routine has been tuned to a site- specific spike 17