Ali Mousavi -- Event modeling
1 like2,580 views
The document discusses an engineering approach to modeling complex systems using event tracking and clustering techniques to improve control and analysis, with a focus on reducing project cycle time by 50%. It compares existing methods like neural networks and machine learning with new event-based methodologies for sensitivity analysis, specifically in real-time applications. The document also outlines a series of steps for event modeling, including tracking, clustering, scenario identification, and validation.
![But What about…
• The Unknowns – The things that we have not noticed but beam data and
information to us. [some NN and Machine learning deal with it but is it
effective/efficient]
• Do they impact my system?
• Is the model sufficient to describe internal and external changes that occur?
• How do I make the system works to specifications when it is in the field?
• Can I change and adapt the system as its internal and external environment
change/evolve?
• Converging with Learning Machine and NN – may be an effective way to
accelerate learning… becoming a member of the family
4](https://image.slidesharecdn.com/eventmodeling-implementation-engineering-140916-engine-160916181356/85/Ali-Mousavi-Event-modeling-4-320.jpg)
Ali Mousavi -- Event modeling
- 1. Event-Modelling An Engineering Solution for Control and Analysis of Complex Systems (SERG Laboratory) Brunel University United Kingdom www.brunel.ac.uk/~emstaam 1Ali Mousavi (SERG)
- 2. Product Engineering, Testing & Validation Process 2
- 3. Existing Methods • Polynomials (Classical DoE): – Classical Linear models, – Mathematically well defined and expressed – Exposed to measurement anomalies – Limited in expressing complex conditions • Neural Networks: – More complex systems can be described – Larger sets of parameters (known Parameters by experts) can be included – Also automatic recognition features (help from Anatoly) – Mathematically beautiful, but implementation and understanding of the solution itself is a challenge – This leads to difficulty to apply calibration – Regular operator and expert interference – over-fitting problem! – Not suitable for real-time applications • Machine Learning – Captures complexity very well – Little need for model calibration and parameterisation – Computationally heavy, rendering it difficult in real-time application – Reliance on statistical analysis throughout (some may consider it strength – some weakness) A new book that I am reading in the subject area: http://www.deeplearningbook.org/ (MIT Press) 3 Deals with Known - Parameters But finds patterns
- 4. But What about… • The Unknowns – The things that we have not noticed but beam data and information to us. [some NN and Machine learning deal with it but is it effective/efficient] • Do they impact my system? • Is the model sufficient to describe internal and external changes that occur? • How do I make the system works to specifications when it is in the field? • Can I change and adapt the system as its internal and external environment change/evolve? • Converging with Learning Machine and NN – may be an effective way to accelerate learning… becoming a member of the family 4
- 5. EventTracking and Clustering If we have the ability to capture all the data possible with the sphere of the problem: • Is it possible to explore their impact on the state of the system one by one and group by group? • Is it rationally and computationally possible/feasible? • What will we achieve by it? • Will the results be in good time to help? • Can I validate and verify them quickly? • Can I put it into any good use? 5
- 6. Typical Systems Modelling 6 The aim is to reduce the project cycle time by 50% using the Event- Based modelling
- 7. Step 1 Event Tracking • Real-Time many-to-one correlation analysis in real time. • Provides a good indication of the impact of various events on performance indicators. • Verifies the known relations and finds new unknown relations 𝑦 = 𝑎1 𝑎2 … 𝑎 𝑛 𝑥1 𝑥2 … 𝑥 𝑛 7
- 8. Step 2 Event Clustering – The Scenario Builder • Event Clustering, coincidence matrix of events at given times • Identifies the group of input and output events that coincide 𝑂1 𝑂2 𝑂𝑛 𝑂1 𝑂5 𝑂 𝑚 𝑂3 𝑂7 𝑂𝑜 𝑖1 𝑖2 𝑖 𝑛 𝑖4 𝑖5 𝑖 𝑚 𝑖9 𝑖11 𝑖 𝑜 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 When these groups of outputs Change these group of inputs Change as well T= the time that this Combination has occurred 8
- 9. Step 3 Scenarios identification Scenario 1 Scenario 2 …Scenario 4Scenario 3 T S 9
- 10. Step 3 State Convergence Scenario 1 Scenario 2 … Scenario n • Compare the matrices of state and combine similar scenarios (distance analysis) • Euclidean distance is a good starter • Number of States is less than or equal number of scenarios State 1 State 2 … State n 10
- 11. Step 4 Look up table for system setting • Go back to the actual values of the parameters • Find cases of interest (e.g. best system condition, worse condition, risk, bottleneck, stability, hazard, …) • Create a lookup table. Condition Representing Bottleneck relevant output and inputs, with values and weights 𝑦1 𝑦2 … 𝑦 𝑛 = 𝑥1 𝑥2 … 𝑥 𝑛 𝑤1 𝑤2 … 𝑤 𝑛 The importance/sensitivity Y X Look up table System setting150 0.76 6.92 1 0.02 1720 11
- 12. Step 5 Validation and verification • Setting up experiments on the actual system or the simulator. • Validate if the they are true solutions, following scenarios my apply: – Instantaneous solution (by setting the controllable parameters the expected results are achieved) – solution with no delay (e.g. changing the angle of a gate to control flow) – The settings are implemented but it takes a fixed amount of time to reach solution - solution with delay (e.g. starting the heating but the material reaches the ideal temperature in t time) – Multiple occurrence – a specific state/scenario needs to be repeated several times for an output to be reached - repetition of a setting for a finite number of time (e.g. failures or breakdown) – Conditional Process – a number of various setting aligned together in a sequence – Process-based solution, scenario 1 to n should align in a sequence for an output to be reached (e.g. completion of an assembly) – Markovian chain 12
- 13. Output EventRelated Input Event (RIE) 𝑡 ≈ 0 Instantaneous Deterministic Event (Current Event Tracker Model) 𝑡 𝑥 𝑡 𝑛𝑡 𝑛−1 … Fixed Time Delayed Deterministic Event – Output event occurs after a fixed time when The relevant input occurs. This can also help us optimise scan rates. Input (𝑡 𝑥) 𝑇=𝑡 𝑂𝑢𝑡𝑝𝑢𝑡(𝑥) Non related Input Events 𝑌 search 𝑡 𝑥 𝑡 𝑛𝑡 𝑛−1… Deterministic singular input multiple occurrence delay. In this case the same input event should repeat itself a number of times until the output event occurs. 𝑛 × 𝐼𝑛𝑝𝑢𝑡(𝑥) 𝑇=𝑡 𝑂𝑢𝑡𝑝𝑢𝑡(𝑥) 𝑡 𝑥 𝑡 𝑛𝑡 𝑛−1… Deterministic sequence of different input events causing an output event (Deterministic Process). In this case a number of specific input event series results in a specific output. Conditional Chain 𝐼𝑛𝑝𝑢𝑡 𝑥 ˄ 𝐼𝑛𝑝𝑢𝑡(𝑦) ˄ 𝐼𝑛𝑝𝑢𝑡(𝑧) ˄ … 𝑇=𝑡 𝑂𝑢𝑡𝑝𝑢𝑡(𝑥) 𝑡 𝑥 𝑡 𝑛𝑡 𝑛−1 … Search in various scenarios for common input event and find the alternative pathways, akin to a tree Petri-Net, Monte-Carlo Tree, …. In this case a sequence of alternative event in a pathway which proceed a common prior event to achieve the final output. The conditional probability of event x occurring at time 𝑡 𝑥 and possible alternative pathways with their consecutive input events to reach Output Y at 𝑡 𝑛. This could be used for predictive or probabilistic pathways to output events 𝑌 𝑌 𝑌 𝑌 𝑡 𝑛: time the Output Occurs 13
- 14. Comparisons between other known techniques • We are yet to stabilise and understand what we are trying to do – with limited resources, challenging well established techniques requires time and collaboration with colleagues • Neural Network – It is underway in a project with JEV Power Plant in Malaysia to explore the optimisation of Harmonic Filters… we will soon release the outcome. • Entropy Based Sensitivity Analysis – small comparisons through a PG dissertation. 14
- 15. EventTracker Verses Entropy Based Sensitivity Analysis Flight Control • Tera bytes of aircraft flight Data • EventTracker algorithm Implementation 15 Data Queuing System Trigger / Event Detection Two way Matching Score Sensitivity Indexes Summation Normalisation Sensitivity Analysis Array Generation Search Slot Analysis Span
- 16. Data Considered Input Data Output Data Aircraft velocity down (veld_gin) Altitude (alt_gin) Aircraft velocity east (vele_gin) Pitch angle (ptch_gin) Aircraft velocity north (veln_gin) Roll angle (roll_gin) 16 These data were considered due to their the apparent relationship between the aircraft velocity and the altitude, pitch angle and roll angle. (simplified for proof of concept)
- 17. Why Entropy • We had worked on the theory before. • Krzykacz-Hausmann1 uses Shannon’s Theory2 of communication to measure uncertainty by. • The entropy is defined as the measurement of loss of information in a system3 . • In the particular case of SIL and HIL, Entropy Based Method can be used as a mechanism for Sensitivity Analysis (Variable Selection) to compare the expected entropy and the actual and possibly interpreted as (sensitivity index). • An important characteristic about the Entropy Based Method is that it is computationally efficient and may be deployed in Real-Time Applications. • As a Sampling based methodology, it requires a sampling generation to decide a verifiable number of observations. • Krzykacz-Hausmann equations for the Sensitivity Analysis were used in this example 1. B. Krzykacz-Hausmann, "Epistemic sensitivity analysis based on the concept of entropy," Proceedings of SAMO, pp. 31-35, 2001. 2. C. E. Shannon, "A Mathematical Theory of Communication," The Bell System Technical Journal, vol. 27, pp. 379–423, 623–656, 1948. 3. B. A. &. B. Iooss, "Global sensitivity analysis based on entropy," Safety, Reliability and Risk Analysis: Theory, Methods and Applications, pp. 2107-2115, 2009. 17
- 18. Comparison at 90msec intervals 18 ptch_gin alt_gin roll_gin 0.19 0.623 0.565 0.347 0.712 0.17 0.347 0.569 0.672 EVENT TRACKER SENSITIVITY INDEX veld_gin vele_gin veln_gin ptch_gin alt_gin roll_gin 0.015 6.773 4.91 0.024 7.1726 0.0070.0155 7.89 5.26 ENTROPY-BASED SENSITIVITY INDEX veld_gin vele_gin veln_gin Range 0-100 Range 0-10 Area of interest… Quite similar results…
- 19. 19 Performance indicator (Sampling time=500ms) Value Number of slots 26 Execution time 253 seconds Memory usage 317.10 KB Performance indicator (Sampling time=90ms) Value Number of slots 15 Execution time 47 seconds Memory usage 317.19 KB Table : Event-tracker performance for 90ms sampling time Table: EventTracker performance for 500ms sampling time Performance indicator (Sampling time=90ms) Value Number of samples 1000 Execution time 101 seconds Memory usage 321.58 KB Performance indicator (Sampling time=500ms) Value Number of samples 1000 Execution time 450 seconds Memory usage 321.58 KB Table: Entropy-based results for 500ms sampling time Table: Entropy-based results for 90ms sampling Longer interval a burst of execution Shorter Interval Significant reduction in execution time
- 20. EventTracker Verses Entropy • Reaching similar results Time could be much more critical factor than memory 20 0 50 100 150 200 250 300 350 Event-tracker Entropy-based Algorithm Performance Time consumption (ms) Memory usage (KB) Performance Comparison Time and Memory
- 21. Various applications • Systems Modelling and Simulation of product Design • Aviation Industry • Manufacturing and Process Engineering • Image Processing • Automotive Industry • Energy and Environmental Studies • Automation • Big Data Analytics 21
- 22. Current Grants 1. Zero Defect Manufacturing €6.2M EU-H2020- FoF – first major investment on the approach – for Adaptive Control of Production Machines and Robots – Commencing November 2016. 2. Previously QMU applied the technique in Deep Drilling EU. 3. Recent Proposals submitted: – Fast Screening and Scanning System for Airport Security (with Russia and UK) – Department of Transport UK – Optimisation of Dairy Industry operations from table to stable (Farm, Production, Storage, Distribution, Retail, and Customer. (EU) – Environmental Monitoring and Risk analysis of Fracking Industry (EU) – Integration of Patient processing hospital and home care (saving an estimate of £5B a year on patient care in the UK NHS – (UK) 4. Working Application: – Efficient Product Engineering, Testing & Validation Process (EPET) in Aerospace industry (UK) – Event-Based motion simulator for people with limb amputation (NATO) 22
- 23. Some sample relevant references from SERG 23 1. Tavakoli S. and Mousavi A. and Peter Broomhead (2013), “Event Tracking for Real-Time Unaware Sensitivity Analysis (EvenTracker)”, IEEE Trans. On Knowledge and Data Engineering, Vol:25, No. 2, pp. 348-359, DOI: 10.1109/TKDE.2011.240. 2. Tavakoli S. and Mousavi A. and Peter Broomhead (2013), “Event Tracking for Real-Time Unaware Sensitivity Analysis (EvenTracker)”, IEEE Trans. On Knowledge and Data Engineering, Vol:25, No. 2, pp. 348-359, DOI: 10.1109/TKDE.2011.240. 3. Mousavi A., and Siervo, H.A. (2016), Automatic Translation of Plant Data into Management Performance Metrics: A Case for Real-Time and Predictive Production Control, International Journal of Production Research – in print. 4. Danishvar, M. Mousavi, A. and Broomhead P. (2016), Modelling the Eco-System of Causality: The Real-Time Unaware Event-Data Clustering (EventiC), submitted to IEEE Trans Systems, Man and Cybernetics – in print. 5. Komashie A. and Mousavi, A., Clarkson, J., and Young T. (2016), A Robust Healthcare Quality Index (HQI) Based on the Generalized Maximum Entropy Formulation, submitted to European Journal of Operational Research – under review.
- 24. Appendix 1. Accelerometer 2. Acceleration along the aircraft vertical axis (acld_gin) 3. Acceleration along the aircraft longitudinal axis (aclf_gin) 4. Acceleration along the aircraft transverse axis (acls_gin) 5. Altitude (alt_gin) 6. Angle of attack from the turbulence probe (aoa) 7. Angle of sideslip from the turbulence probe (aoss) 8. Groundspeed (gspd_gin) 9. Heading (hdg_gin) 10. Rate of change of GIN heading (hdgr_gin) 11. Indicated air speed from the aircraft RVSM (air data) system (ias_rvsm) 12. Latitude (lat_gin) 13. Longitude (lon_gin) 14. Rate of change pitch angle (pitr_gin) 15. Static pressure from the aircraft RVSM (air data) system (ps_rvsm) 16. Pitch angle (ptch_gin) 17. Roll angle (roll_gin) 18. Rate of change of rall angle (rolr_gin) 19. True air speed from the aircraft RVSM (air data) system and deiced temperature (tas_rvsm) 20. Track angle (trck_gin) 21. Eastward wind component from turbulence probe and GIN (u_c) 22. Northward wind component from turbulence probe and GIN (v_c) 23. Aircraft velocity down (veld_gin) 24. Aircraft velocity east (vele_gin) 25. Aircraft velocity north (veln_gin) 26. Vertical wind component from turbulence probe and GIN (w_c) 27. Timestamp 24