Meta-learning for evaluation and implementation of predictive maintenance solution
| Title | Meta-learning for evaluation and implementation of predictive maintenance solution |
|---|---|
| Summary | Finding the best strategy to schedule a predictive maintenance intervention optimizing the cost of unexpected breakdowns against unuseful visits to the workshop. |
| Keywords | Predictive maintenance, metalearning, classification, regression. |
| TimeFrame | Winter 2018 / Summer 2019 |
| References | [[References::[1] Rune Prytz. Machine learning methods for vehicle predictive maintenance using o-board and on-board data. Licentiate thesis, Halmstad University Press, 2014.
[2] Rune Prytz, Slawomir Nowaczyk, Thorsteinn Rögnvaldsson, and Stefan Byttner. Predicting the need for vehicle compressor repairs using maintenance records and logged vehicle data. Engineering applications of articial intelligence, 41:139{150, 2015.]] |
| Prerequisites | Good knowledge of applied data science: classification and regression. Basic knowledge of optimization. |
| Author | |
| Supervisor | Pablo del Moral, Sławomir Nowaczyk |
| Level | Master |
| Status | Open |
Background: there are two main approaches towards predictive maintenance.
- The first one involves classification: looking at historical data of a machine we label as faulty the data happening before a breakdown happened. We will later predict the probability of being faulty, trying to catch a breakdown before it happens.
- The second one involves regression: looking at historical data we label our data with the remaining time to failure. We will later calculate the estimated time to failure.
Usually the work of the data scientists would end here, but there are many more issues arising after our models make their prediction.
When do we send the faulty machine to the workshop? Sending a machine to the workshop or sending a technician to repair the machine has a cost in terms of money and downtime, especially if the machine is not broken.
Should we always send the machine to repair every time our models send an alarm? Should we wait until the next reading and check if the machine is still predicted to be faulty? What if the next reading is regarded as healthy by our classifier? What if the estimated time to failure is small but does not evolve with time? What if the time to failure is big, but is decreasing fast?
There is a big gap between the machine learning models that predict failures and its implementation in the industry. A framework to develop strategies for implementing these solutions and optimization of the cost of unplanned breakdowns and unnecessary repairs is needed for a successful final product.
