Cold start latency is a fundamental challenge in serverless computing. The alternating burst and non-burst phases of function invocations complicate pre-warming, and existing strategies often lack phase-aware modeling, leading to limited effectiveness and high idle resource cost, with poor trade-offs between responsiveness and efficiency. PliKOS is a phase-aware pre-warming strategy designed for pulsed load patterns. It uses I-DBSCAN to identify burst phases from invocation intervals, avoiding full-sequence prediction and reducing overhead. It also adjusts the number of pre-warmed instances based on execution time, request rate, and latency objectives. PliKOS reduces cold start rates by over 95%, lowers idle resource cost by 19%, shortens service time by 13%, and cuts prediction error and overhead by up to 86.4% and 91.4%. By modeling pulsed workloads with phase awareness, PliKOS offers an efficient and practical solution for cold start mitigation in serverless environments.