Abstract: Friction stir welding (FSW) is a multi-physical field coupling process. The acoustic emission signal in the welding process is directly related to the welding defects. Based on acoustic emission detection and multi-feature fusion, the defecting monitored of FSW method is studied. Experiments of prefabricated defect FSW are carried out. The acoustic emission signal in the solid medium is detected in real time, and analyzed by short-time fourier transform, wavelet transform and Mel spectrum which explore the correlation between welding defects and acoustic emission signal. Finally, multi-feature vectors are constructed by the concat fusion method. It is indicated that FSW has different acoustic emission signal characteristics at the prefabricated defects. Short-time fourier and wavelet time-frequency analysis shows that the frequency of acoustic emission signal is concentrated in 20 kHz and the power at prefabricated defects is more than −40 and 0.8 dB respectively. Mel time-frequency analysis shows that the frequency of acoustic emission signal is mainly concentrated in 3.5 kHz and the power is more than −40 dB at prefabricated defects. The multi-layer neural network is applied to establish the welding defect recognition model based on single feature and multi-feature vector respectively. The average recognition accuracy of the multi-feature welding defect recognition model is 97% in the dataset, which is 18% higher than the single-feature defect recognition model. The multi-feature welding defect recognition model can more accurately recognize and monitor the welding state.