Best algorithm for predictive maintenance / please suggest
I need a suggestion from you, the experienced ones, regarding best algorithm for anomaly detection.
Use case: there are six bearings on a single device. It use to happen a bearing slowly stop functioning correctly until it gets stuck and the device stop working. To predict this scenario and change the bearing in advance, temperatures are measured for each bearing. The malfunction happens always with one bearing at a time.
There has been already a proof of concept project for this very case, which proved it can predicted. Unfortunately i have no further details/outputs from this PoC. Only what i have is data of measurements (temperatures) for period of two months. Data is not labeled, but when visualizing them in a plot, there can be found examples when one of the temperatures had higher trend and finally started to rise even more. Then came back to normal/average after it was fixed.
Now i have a dilema and that's why i need your suggestions. I want to use a ML approach for this project and i am sure it can be done. On the other hand it could be done by simply calculating the average/median of 6 temperatures and comparing which temperature is too much above.
In the 2 months data i have is obvious there is a relationship between outside temperature and the temperature of bearings. The difference in average temperature at the beginning (November) and end of data (January) is approx. 2 degree Celsius. And that was in winter, in summer difference can be much higher (compared to winter). On the other hand, the relative range of temperatures is approx. same.
So considering the luck of data it can be problematic to have a good anomaly detection algorithm. If i would have data for whole year i would not hesitate to use ML. But maybe it would be better to solve it by calculating the median and checking for outlier. But damn, i want a ML approach!
What is your opinion?
Thanks a lot.