Ultrasonic metal welding is a high-frequency vibration process that creates joints for components and has attracted extensive research attentions in recent years. Although this technology has been applied in many areas such as electronics, automotive, and aerospace, its wide adoption is still hindered by unsatisfactory welding quality. In such cases, performing process monitoring to enhance the welding quality is of great significance. In current literature, the exploration in monitoring and inspection of the ultrasonic welding process is still limited due to the ineffective offline monitoring and the lack of investigation of the anvil state impacts on welding results. To address this problem, an acceleration sensor monitoring system is proposed in this study aiming for enhancing the welding quality. In the established system, the acceleration signals are collected and decomposed by variational modal decomposition (VMD) to capture the characteristics of the intrinsic mode functions (IMFs) energy distributions. In addition, the particle swarm optimization (PSO) algorithm is adopted for penalty term selection while the system frequency and sampling rate are used for mode number determination. The proposed method is experimentally validated and suggests high effectiveness and robustness for anvil state identification in ultrasonic metal welding.