# Machine learning to predict Equipment Failure

I am in a confused state of Mind to actually decide if My use case does qualify to be resolved by Machine learning algorithms.

I have a temperature measuring sensor connected of an equipment. so the sensor mainly captures two attributes , One being temperature itself and other being Time of measurement. My Aim is to send out an alarm signal when the temp remains high after certain time frame , Lets say the temp sores to 70 degree and remain constant or increase even after 1 Minute (Sensor measure data once in every 5 seconds) - So with successive 10 reading shows the Temp not going down and remains As high as 70 degree or sometime going further up - This is a Alarm situation and Must predict that equipment can burn out with such high temperature.

So in this case , the previous Dataset is of no use , as the outcome is predefined (Temperature >= 70 degree & stable for 10 readings) . Can this be treated as Machine learning use-case ? I only have live data and no prior data for this use case.

• are the temperature limits pre-defined? how do you know if should trigger an alarm or not, is this same for every equipment device? – miraculixx Sep 29 '16 at 10:04

If you have an hard rule (Temperature >= 70 degree & stable for 10 readings) , you don't Machine Learning. If you want to learn a better rule to detect failure from temperature alone (maybe associated to the equipment associated to the sensor) you can use machine learning or, in this specific case, time series analysis. There's a lot of stuff already done on the subject, I believe.

When you say:

"...My Aim is to send out an alarm signal when the temp remains high after certain time frame..."

I understand that this situation is process monitoring, not Machine Learning (ML), because you want to just issue an alarm or alert for a given condition for possible known fault mode.

It will be ML supervised if you want to predict the temperature or process failure and or instrument for the case of a regression or classification, considering your objects/observations.

For your problem, perhaps time-series or analytic-based approaches, such as Kalman filter, can also solve your problem.

I suggest that this question be transferred to Data Science, Artificial Intelligence or Signal Processing.