Compare each incoming reading with the previous value. If the new value is less, then the previous reading was a peak.
This method is fooled by one downturn (or noise). It's more reliable if you average a few readings in a row, so as to ignore small downturns.
Using 5 as an example. Take 5 readings in a row. Tally and divide by 5.
Store the result as variable 'X'.
Take another 5 readings. Average them.
Compare the result with 'X'. If the new figure is lower, then branch to your routine that recognizes a peak.
Store the latest figure as the new value for 'X'.
You'll need to experiment to discover how many readings you need to average. The aim is to find the best amount for cancelling the effects of noise.