During laser welding of pure copper, instabilities such as violent molten pool oscillation, large spatters, and melt ejection severely damage weld quality, which are caused by copper’s high reflectivity on commonly used infrared laser. The seriously unstable molten pool and keyhole and complicated laser-material interactions result in complex process signal emission waveforms, adding to the difficulties in process stability monitoring tasks. In this work, to break down different contents in the complex signals and deeply analyzing signal-process relation, combinative spatial optical sensor system was designed, and time-frequency signal analysis in multi-scale windows was performed. It was found that the infrared radiation at the front and end side of molten pool indicates the oscillation behavior of liquid metal surface, and the signal fluctuation patterns of visible radiation from different height of metal vapor varied when meeting severe instability like melt ejections. Signal features were extracted based on the understanding of process mechanism and signal behaviors. A cascade model combining Artificial Neural Network (ANN) and Support Vector Machine (SVM) was introduced to predict weld seam quality, where the ANN model focused on short-time stability status perception and the SVM model was used to decide macroscopic seam formation defects based on combining outputs of the ANN model in a long-term sampling window. Application results showed that the recognition accuracy of pit was 100 % and the accuracy of uneven toe reached 86.3 %. The multi-source signals of unstable molten pool recognized by the cascade model were summarized. The evolution process of copper molten pool ejection was revealed.