Meeting Minutes

10. and last Meeting, Mai 26., 10.30 am

Write "Review" chapter including a numbered list of outcomes and difficulties, further work recommendations.
Write "Litterature review" with emphasis on comparing (final findings + approach details) to the paper of Wu et al. Refere to bibliography in the text.
Include more real applications, data and interpretations

Current status: For a detailed report see here.
- Clustering
- Data Generation
- Finding Optimal Configuration
Start writing the report until next week, focusing on chapters 2 and 3.

Use data generation to overcome problem of few data items, look for references on the subject.
Binary Cost Approach -> Paper
Define a concrete maintenance plan: Values --> Action, Jongho will send other report as reference.
What about quantizing input values? Compare binary vs. continues input values.

Post Meeting with Jongho:

Data generation using K-Means clustering and assume non-correlated, gaussian RV. (only for failure states, there is enough other data).
Do clustering on the whole input data available (ie. not split in test and training) to get better clustering results.

Please refer to the status report of week 6 (here) for more details.

Discussion of problems encountered during dataset preprocessing and first test run.
- Makeing the dataset 'continuous' requires many assumptions to be made. (Not working on timeseries for now but rather on environment conditions)
- Mixing up data of different systems. (Focusing on electrical system and related codes)
- Meaning of Error Codes? (Clear thanks to Error Code List from Jongho)
- --> Continue with one system only and related error codes, use 'windowing' method to extract sample data for times where the system is in a good state.
Dimitries asks for schematic drawing of ANN.
Graph with intersections of global failures with smaller failure events looks interesting --> follow up later on?

Short discussion of proposed TOC and schedule.
Goal for next week should be an evaluation of existing, similar approaches in terms of
- Input data used
- Output type (categories vs. real values, status, cost or decision)
- ANN structure
Propose specifications to be used in this project.

Following extended discussion with Jongho:

How can the existing data be used to be feed into an ANN? -> No continuous time series data available, only at failure instants. But failure != break down, there fore the failure rate combined with environmental data (are they correlated?) could serve as input.
What should the output value of the ANN be? -> The ANN needs to be trained, we need a set of training and test data with "solution".
Traditionally a predictive maintenance approach is split into 2 parts: From measurements to expected MTTF and from there to alternatives, related cost and a decision. Should this project to the 1st part only or both? -> So far mostl the 1st has been tried, to do both would be interesting.
It would also be nice to test against other, existing approaches (eg. statistical analysis, condition based approach, etc.)

Introduction to the project idea.