First of all … you should ALWAYS model TIME SERIES i.e. bucketed data which is observed NOT what is accumulated UNLSS you wish to first bucket/accumulate transactional data to create a bucketed time series. The time series to be analyzed should never be an unneeded accumulation or an unneeded differencing.
The data you posted is here
, When you accumulated your data you injected non-stationarity (trend in this case) into your new series which you posted as a picture.
A useful model for the original data is obtained here containing the answer to your question.
. The evidence suggests that the non-stationarity in your observed data ( starting at 2016/4 ; 48 values ) is as follows:
1) there is a systematic seasonal pulse in December of each year (period 9) caused by an unspecified but latent exogenous deterministic effect possibly anthropogenic in nature.
2) there was a level shift DOWN at or about period 9 (2016/12)
3) there was 1 unusual activity DOWN at period 12 (2017/3)
4) there is significant positive correlation between observations 2 periods apart
I used AUTOBOX , which I have helped to develop, but essentially the analytical tools of Intervention Detection and arima model identification were simultaneously employed.
The residual plot is here
and the acf of the residuals suggesting sufficiency of the model is here 
The Actual/Fit and Forecast graph is here
providing integer forecasts and forecast intervals for the next 12 periods.