Aiming at the reduced inertia level, diversified inertia sources and enhanced dynamic coupling of power systems under high-penetration new energy integration, this paper proposes a multi-source inertia collaborative estimation method based on meta-reinforcement learning. The method integrates the dynamic characteristics of synchronous generator inertia and new energy virtual inertia to establish a unified multi-source inertia state representation model. Considering sparse measurement scenarios, the multi-component inertia estimation problem is formulated as a partially observable Markov decision process, and a composite reward function is constructed with frequency stability indexes to realize unified inertia estimation from time-series measurement data. By integrating PGD adversarial enhancement, A3C asynchronous optimization and MAML meta-transfer learning, a three-layer collaborative Meta-A3C framework is established. Hierarchical gradient optimization realizes noise suppression, time-series decoupling and fast cross-condition adaptation. Simulation results validate that the proposed method improves the accuracy and robustness of inertia identification in complex power grids, which can provide technical support for frequency safety assessment and stable operation of low-inertia power systems.