Safety Enhancement for Deep Reinforcement Learning in Autonomous Separation Assurance