Abstract: Experimental analysis of the performance optimization and fixation methods for large-scale ultrasonic welding heads was conducted. Parametric processing of the long slot dimensions of the welding head was performed, generating 360 sets of randomized sample data and corresponding dynamic performance parameters. An 18-12-8 structured back propagation (BP) neural network model was constructed to establish a mapping relationship between long slot dimensions and dynamic performance, achieving a model prediction error within 10%, maximum longitudinal amplitude error ε₁ ≤2.58%, and maximum stress error ε₄≤7.02%. Subsequently, the multi-objective particle swarm optimization (MOPSO) algorithm was integrated to maximize longitudinal amplitude γ₁, transverse amplitude γ₂, vertical amplitude γ₃, and stress σ_max, resulting in a Pareto optimal solution set. The optimized welding head exhibited a 30.43% increase in the maximum longitudinal amplitude from 23.0 μm to 30.154 μm, with stress distribution meeting the material requirements of the welding head. Based on the analysis, a novel fixation method was designed by installing support tools at the welding head’s side with low vibrations. The results show that the stiffness of the optimized fixation scheme is improved, with remarkable vibration isolation effects, and the maximum longitudinal amplitude difference of the welding head is only 0.467 μm. The study validates the effectiveness of the MOPSO-BP hybrid optimization strategy, providing technical support for enhancing the performance of industrial ultrasonic welding systems.