Cellular IoT devices cannot stay connected to AWS

Reconnections, or TCP session timeouts, via mobile network operators (MNOs) can be challenging. Most use some for of carrier grade NAT that will manage connections from the mobile device to the public Internet. You can get a better indication of which side (client or IoT server) disconnected from the CloudWatch logs.

One thing to test is to reduce the device's MQTT keepalive to a lower value and monitor. While each MNO is different, and can be different depending where the are deployed, I would start with a value of 270-290 seconds, which is just under 5 minutes--a common value seen for expiring inactive TCP sessions.