Predicting Building Fire Alarms

machine_learning
arima

#1

Hello,

I have Building fire-alarm data for around 6 months - Date and the no. of building fire alarms triggered for that Date. I have to forecast the no. of fire alarms that can trigger in future based on this past data.

I treated this a time series problem and used ARIMA for forecasting. I got the forecasted values but I am not convinced with the output and the strategy. I am thinking, whether it makes sense to do it? No. of fire alarms are not seasonal etc. They may depend of values from various other IoT sensors. Does it make sense to forecast no. of expected alarms based on past data? Should I try to get more (more features) data or my existing strategy is fine to go ahead with? Please suggest what makes most sense to do in this type of problem. How to frame this problem? What type of data should I collect and how should I go about solving this problem.

Please help me get some direction.

Regards,
Nandy.


#2

Can you share few records of your data ?


#3

Hello Paragdgu,

Data looks like this…

Date
2018-06-01 1.0
2018-06-02 1.0
2018-06-03 4.0
2018-06-04 0.0
2018-06-05 4.0
2018-06-06 1.0
2018-06-07 1.0
2018-06-08 1.0
2018-06-09 0.0
2018-06-10 5.0
2018-06-11 1.0
2018-06-12 4.0
2018-06-13 1.0
2018-06-14 2.0
2018-06-15 4.0
2018-06-16 1.0
2018-06-17 1.0
2018-06-18 0.0
2018-06-19 2.0
2018-06-20 4.0
2018-06-21 0.0
2018-06-22 2.0
2018-06-23 2.0
2018-06-24 1.0
2018-06-25 0.0
2018-06-26 2.0
2018-06-27 4.0
2018-06-28 0.0
2018-06-29 2.0
2018-06-30 1.0

2018-10-07 1.0
2018-10-08 0.0
2018-10-09 3.0
2018-10-10 1.0


#4

@nandita38

Model 1:I think you can model it using poisson distribution. You can calculate the mean number of alarms per day from the historic data. This will be your lambda. Now you can calculate the probability of occurence of x alams onn any given day.

Model 2:An improvement on this will be creating a new column which is week number of a month.It can have values from 1 through 5. Now calculate the mean number of alarms for week 1 and repeat Steps in Model 1. In this way you can calculate the probability of x alarms in week 1 for any month. You can repeat the steps to get a poissson distribution of number of alarms for Week 2 through Week 5 .If you had more data you could have gotten to a daily level instead of weekly


#5

Hi @paragdgu Parag,

Thank you for your inputs. It totally makes sense. I have now got some more features for this data i.e more data regarding it. Will it be possible for you to take a look at it and suggest what could be the revised strategy, in your view.

Regards,
Nandita.


#6

Yes please share the data in csv format