ANALYSIS AND VISUALIZATION OF TIME-VARYING DATA USING ‘ACTIVITY MODELING’ By Salil R. Akerkar Advisor Dr Bernard P. Zeigler ACIMS LAB (University of Arizona)
Presentation Outline Introduction Activity – A DEVS Concept Activity Modeler System Stage1 - Preprocessing Stage2 - Activity Engine Stage3 - Visualization Results Implications for Discrete Event Simulation Future Work
Introduction Data Source and Problem under study Current trends Unexplored area Motivation – Discrete Events vs. Discrete Time
Activity – A DEVS Concept Definition of Activity t1t1 0 titi m1m1 mimi mnmn T
Activity – A DEVS Concept Coherency (Space and Time) Instantaneous Activity Accumulated Activity (same as DEVS Activity) Activity Domain
Activity Modeler System Raw Data Raw Data FORMATTED DATA RESULT S Stage-1 RESULT S ACTIVITY DATA ACTIVITY DATA Stage-2Stage-3 PERL FORMATTER ACTIVITY ENGINE (OPTIONAL) PERL FORMATTER GNUPLOT MODULES GNUPLOT MODULES AVS- EXPRESS MODULES
Stage 1 – Pre Processing Why do we need pre-processing? Regular Structure format PERL formatter Functions Extract Information Format Correction Logic Analyze part of information 2D formatter decrease IO operations standardization
Stage 2 – Activity Engine ACTIVITY GENERATOR PATTERN PREDICTOR STATISTIC ANALYZER ACTIVITY TIME- SERVICES ACTIVITY LOG THE ACTIVITY ENGINE PERL Formatter GNUPLOT SCRIPTS AVS-EXPRESS MODULES DATA-FILE PATTERN INFORMATION STATISTICAL INFORMATION ACTIVITY DATA DATA ENGINE
Stage 2 – Data Engine Functions File handling Sequential / Random access Standardization of filenames for automation Memory Allocation Transformation between domains Cellular and Temporal Transformation between dimensions Val2D[i][j] = Val1D[i*Cols+j] Independent of spatial dimension
Stage 2 - Activity Generator Instantaneous Activity Accumulated Activity Time Advances Activity Factor (AF) Cellular domain Threshold (AF) Cells Activity factor
Stage 2 – Statistic Analyzer Extract Statistics in terms of groups Group1: Maximum, Minimum, Range, Average Group2: Standard deviation, Mean Group3: Living Factor (Temporal domain) Group4: Histogram of Time Advances Static in nature Provides meaningful threshold to Activity Factor Living Factor
Stage 2 – Statistic Analyzer Group 3: Living Factor (LF) Temporal domain Group 4: Histogram of Time Advances Temporal domain Logarithmic in scale Time
Stage 3 – Pattern Predictor Spatial and Temporal Coherency Peaks and Max Analyze activity pattern Predict activity pattern
Stage 3 – Pattern Predictor Max Locator Peak Locator Difference in Peak and Max False Peak problem Eliminated by ROI (Region of Imminence)
Stage 3 – Region Of Imminence (ROI) Definition Steps Peak Detection in IA Scanning algorithm Boundary conditions Threshold conditions ( ) Significance Imminence Factor Cells
Stage 3 – Pattern Predictor 1D scanning algorithm 2 neighbors Binary visualization Cells Threshold condition Boundary condition Peak Under consideration: 2 Location of cell: 10 Initial Values: Left-neighbor = right- neighbor = 10 Final Values: Left-neighbor = 7 Right-neighbor = 13
Stage 3 – Sphere Of Imminence 2D scanning algorithm 3 types of tuning Coarse Normal Fine
Stage 3 – Sphere Of Imminence Type Of Tuning Computation time (ms) Imminence Factor (t= 5) Imminence Factor (t=10) Coarse Normal Fine Coarse Tuning Normal Tuning Fine Tuning
Stage 3 – Region Of Imminence ROI: Overcome the False Peak problem
Stage 3 – Predict Pattern 1D space Linear Span Module [0.9 – 0.95] Order of Pattern Pattern attributes Offset Direction Difference Steps Recognizing pattern t[n,n+1] 5 1 st order pattern 2 2 nd order Predicting pattern t[n+2,T] t = ta t = tb t = ta t = tb ROI Linear span 2 nd Order 1 st
Stage 3 - Visualization Softwares GNUPLOT AVS-Express Visualization Stages Reader (Import data) Visualization modules Writing stage Reader VIZ modules Writer
Stage 3 - Visualization Zero Padding Binary Visualization Advantages Eliminating unwanted data Reduction in file size Implementation set zrange [0.5:]
Stage 3 - Visualization DomainTypes Of ResultVisualization Techniques 1D 2D Spatio-Temporal Instantaneous Activity, Accumulated Activity, Time Advances Surface Plot Images (GNUPLOT) Surface Plot / Contour movies (GNUPLOT scripts/ AVS-Express) Region Of Imminence, Peak Locator, Max Locator Binary Visualization, Zero Padding (GNUPLOT) Binary Visualization, Zero Padding (GNUPLOT scripts) Cellular Statistics, Activity Factor 1D single / multi graphs (GNUPLOT) Surface Plot Images (GNUPLOT) Temporal Living Factor, Histogram of time advances 1D single / multi graphs (GNUPLOT) Surface Plot Images (GNUPLOT)
Results 1D space 1D heat diffusion process Robot Activity 2D space 2D heat diffusion process Fire-Front model
Results – 1D Heat diffusion 1D space,T=100 N=10, 100, 200 N Cells Time
Results – Robot Activity 1D space Robots modeled as cells Simulation time steps – 2357 Data (Value domain) 1- Robot moving 0- Robot stopped Activity domain 1- State transition 0- Same state Robots Time
Results – Robot Activity Living Factor Activity Factor Imminent groups
Results – 2D diffusion 2D space (100 x 100 cells) T = 50 Cellular domain results (2D) Activity Factor Statistics Surface plot images IA surface characterized by concentric circles t adv histogram lower end Activity Factor Histogram of Time Advances
Results – 2D diffusion Movie of IA / AA (activity domain) and output values (value domain)
Results – Fire Front model 2D space (100 x 100 cells) T = 297 Movie for Value domain
Results – Fire Front model Living Factor 20% maximum t=180 boundary Imminence Factor = 0.7 t [50-150] Time
Results – Fire Front model Instantaneous Activity Peak Bars Accumulated Activity Region Of Imminence
Implications for Discrete Event Simulation DEVS transitions: DTSS transitions: Maximum Slope: DEVS v/s DTSS
Implications for Discrete Event Simulation MODELCELLSTIMEMAX(IA)TOTAL AADEVS DTSS 1D diffusion (N=10) D diffusion (N=100) D diffusion (N=200) D diffusion Fire Front DEVS v/s DTSS
Results – Predict Pattern Test data - 3 1D diffusion (N=100)
Results Results for 1D process Test data 1D diffusion Percentage Error decreases as N increases ROI characterized by linear curves
Conclusion New perspective for data analysis – Activity domain ROI – Spatial Coherency in Temporal domain Analyze process behavior in terms of Activity Compute and Predict – activity pattern Results – process specific Predict Pattern - % Error decreases as N increases ROI curves are characterized by linear curves DEVS found to be more efficient than DTSS
Future Work Extending system to data in 3D space Extending system to UNIX platform Enhancing the Pattern predictor module Efficiently Detecting the ‘new Imminent Cells’ in DEVS simulation
ACKNOWLEDGEMENTS Dr. Bernard Zeigler Dr. Salim Hariri Dr. James Nutaro Dr. Xiaolin Hu, Alex Muzy Hans-Berhard Broeker Cristina Siegerist ACIMS LAB
QUESTIONS ?