KPI Degradation/Network Failure Predicition!

r
classification
predictive_model
forecasting
failure

#1

Hi All,

Here is a challenging puzzle which has been blowing my mind for last couple of days , hope it will do same magic for you also :smile:

Domain:- Cellular Antennas
BTS-Base station ,which has multiple cells
RNC-controls multiple BTSs.

Provided Data has some KPIs like Access failure and Drop Rates and counters which were recorded on hourly basis. This Data is at cell level.No regional or physical condition being provided with it.

Objective of the case study is to predict when will the KPI breaches the threshold value for an hour for a particular cell.

Hurdles:-
1.For almost all of the cells Variables including Dependents(KPIs) have 99% zeros.
2.Every cell ,every BTS behaves differently.hence unable to figure out generalize seasonality. as there are thousands of BTS and cells are 7-8 times for a bts. didnt find any pattern to segment also
3. cant go with linear regression as everything is zero. nor with Time series , as data point for each cell is 2000 unable to figure out any trend or seasonality.

Can anybody help if gone through similar kind of problem.

Thanks in Advance!


#3

i don’t have knowledge in cellular domain
do you have any related predictor variable provided?


#4

Perhaps you try to find a rolling time and then measure the variance and then work on the outliers. As it is a signal or I understand like this you could have a lot of noise and then the mean does not mean anything, therefore linear regression goes out.
So you should try non parametric model, quantile regression perhaps. I have no clue about the data therefore it is difficult to advise
Hope still it helps.
Alain