The fact that you are seeing CEA with Success means that the peer ends have agreed upon the Application ids and the VS attribs. Moreover you are even seeing the DWR being generated at PCRF.
I think somehow your DWR is not being sent out therefore the PCEF is timing out and RSTing the connection.
May sound wierd, but can you check your IP route table to see if there are ambiguous routes (especially through different interfaces) to the PCEF? There have been such instances where some requests/keepalives are sent, but take a different path and never reach the other end, thereby flaking up and down the connections.
Another approach would be, use a simulator (if you readily have one) to simulate your endpoint(depends on whom you suspect, the PCEF or PCRF) and see if the same behaviour occurs.
Run continous ping from one end to another and see if there are any packet losses.