That means you have a persistence issue. Either saving the nonce or retrieving it after restart fails.
What is the interval between join attempts? If you choose it too small (less then a couple of minutes after the first two attempts with a backoff mechanism that increases on every failed attempt) you are creating a new issue because the gateway will have too many downlinks to send and will drop some of them resulting in even more join failures. Hammering the network with requests never works.
Both. It is a security feature of LoRaWAN and a bug in the end node software. Sorry.