Divergence after adding manual features before fully-connected layer

I’m training a ResNet, which could classify ECG signals.The input is 1-D signals containing 300 sample points. When I use just ResNet, the results are convergenece. Then I try to add manual features before FC, concatenanting 256 ResNet features and 14 manual features. The result is divergence.How can I solve this problem? Should I modify manual features’ order of magnitude?
This my structure: