![Page 1: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/1.jpg)
An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform
Behrad Bagheri
Linxia Liao
![Page 2: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/2.jpg)
2
► Linxia Liao» B. Sc. In Mechanical Science & Engineering, 2001, HUST, Wuhn,
China» M. Sc. Mechanical Science & Engineering, 2004, Huazhong
University of S&T. » Ph.D. Mechanical Engineering, 2010, University of Cincinnati
» Internship at Harley-Davidson Motor Company» Visiting Scholar at Siemens Corporate Research» Research scientist at Siemens Corporate Research
About the Author
![Page 3: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/3.jpg)
3
1. Introduction2. State of the art3. Degradation Status Assessment4. A Framework for Prediction Model Selection Based on Reinforcement
Learning5. A Novel Density Estimation Method to Improve the Accuracy of
Confidence Value Calculation6. Design of a Reconfigurable Prognostics Platform (RPP)7. Conclusion and Future Work
Outline
28 March 2013
![Page 4: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/4.jpg)
4
► Assumptions» Certain Vibration Signals can indicate the health of a system» A confidence value threshold can be set to indicate acceptable
performance or a serious failure» The system being monitored is degrading gradually in an
observable and measurable way.» The baseline is consistent for a certain period of time
Assumptions and Challenges
![Page 5: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/5.jpg)
5
Degradation Status Assessment
![Page 6: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/6.jpg)
6
► Feature Extraction from Vibration Signals
Degradation Status Assessment
► Dimension Reduction -> PCA
► Evaluate Degradation Status by SOM» MQE Health Assessment
![Page 7: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/7.jpg)
7
► Experiment Configuration» Two ICP Accelerometers for each bearing» Sampling Frequency 20 kHz, Sampled every 10 minutes for 2 seconds» A magnetic plug in the oil, used as evidence of system
degradation(Amount of debris on the magnetic plug increases when bearing wore out)
► Feature Extraction (11 Features)► Dimension Reduction
» Top two principal components with 90% of variance
Case Study – Bearing Run-to-Failure
![Page 8: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/8.jpg)
8
► SOM-MQE Degradation Status Assessment» First 500 cycles used as baseline data» 4 sections could be distinguished in the MQE plot
Case Study – Bearing Run-to-Failure
![Page 9: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/9.jpg)
9
A Framework for Prediction Model Selection Based on Reinforcement Learning
![Page 10: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/10.jpg)
10
» Adaptively choose the best prediction model for predicting the feature for each step
Description of the Concept
![Page 11: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/11.jpg)
11
► Elements of Reinforcement Learning» Environment: Historical data from database. » Action: the ARMA model used for prediction» State: different degradation states determined by MQE values» State Transition » Reward: A function related to prediction accuracy.
Elements of Proposed Method
![Page 12: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/12.jpg)
12
► Reinforcement Learning trains an agent to interact with the dynamic Environment
► The target is to maximize reward in a long run of trial and errors► Look-up table created by Q-Values is used to select models
![Page 13: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/13.jpg)
13
Case Study – Bearing Run-to-Failure
► 6 ARMA models and 1 Linear model are used for prediction► 9 States, prediction for 20 Steps
![Page 14: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/14.jpg)
14
► Using the results of 3 runs is more reasonable in selecting model
► In case that for the same state more than one model have the same probability, Occam’s razor principle could which states the simplest model should be selected
![Page 15: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/15.jpg)
15
► First principle component of input data is used for prediction. ► 3rd run is used for training (Environment) and 11th run is used for
testing► 10 states are defined in one run along with 4 ARMA models.
Second Case Study - Spindle
![Page 16: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/16.jpg)
16
Case Study 2 - Results
![Page 17: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/17.jpg)
17
A Novel Density Estimation Method to Improve the Accuracy of Confidence Value Calculation
![Page 18: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/18.jpg)
18
► Study the distribution of predicted features and comparison with the distribution of baseline data will result in calculating CV value.
► Boosting Algorithm of Gaussian Mixture Model (GMM) » PSO is used to optimize the selection of Gaussian models
Calculation of CV – Boosting Algorithm for GMM
T: Number of Mixturesx: training datasetαn: coefficient for each h(x)h(x): weak learner
![Page 19: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/19.jpg)
19
Case Study – Bearing Run-to-Failure
► DLL value for Boosting GMM shows that this algorithm has better performance than two other methods
► Feature values for next 20 steps are predicted using the Boosted GMM, GMM with PSO and GMM Only methods
► Red dots show the predicted values, black and purple dots show high and low 95% confidence boundaries
![Page 20: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/20.jpg)
20
Design of a Reconfigurable Prognostics Platform (RPP)
![Page 21: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/21.jpg)
21
Reconfigurable Prognostics Platform (RPP)
SA: System AgentKA: Knowledge AgentEA: Executive Agent
![Page 22: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/22.jpg)
22
► Two case studies for RPP evaluation
ATC Health Monitoring Spindle Bearing Health Monitoring
![Page 23: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/23.jpg)
23
Evaluating RPP with Case Studies
► Steps and related spent times in reconfiguring server for new request
![Page 24: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/24.jpg)
24
► SOM MQE method can provide a quantitative measure of the machine degradation with only baseline data
► The reinforcement learning framework utilized ARMA models as local prediction agents. The proposed method selects appropriate prediction model to gain better prediction accuracy
► The proposed density boosting method to convert prediction results of the feature space into confidence value yields more accurate estimation of CV Value
Conclusion and Future WorkConclusion
Future Work
► Identifying the critical components of the complex systems.► Considering more signal processing methods to prepare raw signals► Platform synchronization & standardization
![Page 25: An Adaptive Modeling for Robust Prognostics on a Reconfigurable Platform](https://reader036.vdocuments.us/reader036/viewer/2022062323/56815b1f550346895dc8d488/html5/thumbnails/25.jpg)
25
Thank You