In order to classify typical penetration states in laser welding process, multi-sensors are applied to acquire the audible sound, ultraviolet and infrared (IR) emission signals, also penetration-state-relating features of each signal are extracted. Based on feature-level information fusion, simulated annealing algorithm is utilized to optimize charactristic signals, consequently setting the “coeffients of feature fusion”. Pattern classifier is designed using back propagation (BP) network. It is found that through samples training and optimizing, a classification of 88%～100% has been made for detection of the four distinct penetration states such as “excessive penetration”, “full penetration”, “unstable penetration”, and “partial penetration”. So, an effective method for on-line monitoring for laser welding quality is provided.