Making Factories Smarter Through Machine Learning
This failure happened to a production line with a turnover capability of 30 M € per day with 50 operators directly working on the line per shift. When the failure occurred, the complete line had to be stopped during 3 shifts until the repair was completed. However, these failures may last for at least 30 shifts and the repair time is directly related to the availability of spare parts. In this case, the workforce cost and the lost turnover should be added to the complete cost.
It is important to state that a machine learning-based monitoring system could detect the first failure peak, giving enough time to stop the line in a controlled manner. A production and workforce could then be reassigned to reduce the failure impact over production line productivity. In this case, a relatively simple machine learning algorithm is able to detect these types of anomalies within a variable, effectively warning about the problem when a first peak is detected or before.
Figure 3: Summary of costs for damages due to bearing malfunction
4. MACHINE LEARNING FOR PREDICTIVE MAINTENANCE
Fast forward a few years to a new system where machine learning is being used to analyze the data to predict potential system failure 14. First an understanding of the machine learning approach is needed to gain full appreciation of the scope of the predictive analytics performed to predict when the component / system may fail and even more importantly why the failure may occur. This, in turn, allows system optimizations that can extend the lifetime of the asset and the overall system.
14
D. GOYAL y B. S. PABLA, « Condition based maintenance of machine tools— A review,» CIRP Journal of Manufacturing Science and Technology, vol. 10, pp. 24-35, 2015
- 34- January 2017