Here is a way. Specifically, this looks back 1h, and returns the number of minutes the instance you care about has been down.
((240 - sum_over_time(probe_success{instance="192.168.1.71"}[1h])) * 15) / 60
In the above case, prometheus is polling every 15s, so every hour has 3600 / 15 = 240 samples. We add up all the 1s returned by the probe_success metric, take difference from 240, and multiple by the 15s polling time. Then divide by 60 to get minutes. Obviously adjust the math to your desired "look back" interval and polling rate.
It's not perfect because if your instance flapped a bunch of times in that hour, we are still computing total downtime, not time since last time it went down and stayed down, but should be good enough.