Looks like one of the machines has a bad view of the cluster:
184.108.40.206:1099, 220.127.116.11:1099, 18.104.22.168:1099
The explanation is as follows:
Each JGroups node has an address consisting of IP address and port, e.g. hostA:12345.
A JGroups address is mapped to a JBoss address, which looks like hostA:1099
When a JGroups node bounces, it'll get a new address, e.g. hostA:23456, however the JBoss address is still hostA:1099.
So when hostA bounces, and the bounce is not detected, e.g. because your timeouts in FD are high, then we might have a JGroups membership of , which is translated to a JBoss membership of . This is the case until the bounced former member is removed (hostA:12345).
If you set the timeouts in FD to a lower value, or use FD_SOCK, this will go away