The lock-loss issue has been fixed! 

The lock loss happened not in the IMC but in the FSS, and the actual bad guy is the AOM driver AM input. As reported in klog6755, I put a filter bank for the channel to input DC signal into the AOM AM, K1:PSL-AOM_AM. I did not connect the input of this filter bank to anywhere, and it might be a problem. I could not imagine the reason, but the mysterious unknown signal was injected into this filter bank. This signal increased 1 count in one minute, and once it reaches to the range of the digital system, the FSS loop lost lock. The attached plot shows the output of this filter and the REFCAV-TRANS. 
I modified the model and terminated the input of this filter. 

We hope the input optics continue to lock eternally.