Abstract: Chatter induced by the regenerative effect in milling can severely deteriorate machining quality and efficiency, and the online identification of chatter is critical for ensuring stable machining. To address the problem that chatter characteristics vary significantly across operating conditions under variable tool posture milling, which renders it difficult for single-representation methods to accurately identify chatter, a multi-channel thin-walled part milling dataset covering different inclination-angle conditions was constructed, and a chatter identification method based on time-frequency and modal dual-domain fusion was proposed. Multi-synchrosqueezing transform (MSST) and successive variational mode decomposition (SVMD) were employed to extract time-frequency image samples and modal component samples, respectively, so that the machining state could be characterized from two perspectives, namely local time-frequency energy evolution and overall modal structure. On this basis, a time-frequency branch, a modal branch, and a dual-domain fusion network were established to distinguish stable and chatter states. Experimental results show that the proposed method achieves a test accuracy of 96.82%, outperforming both single-branch models and other comparative methods overall.