Monitoring of weld joint quality is a significant issue in Cold Metal Transfer (CMT) lap welding. In this paper, CMT lap welding experiment of low carbon steel sheet was carried out, the sound characteristics of CMT are studied. Further analysis of the “special two-step mode” of welding sound shows that the faster the change of arc energy, the greater the corresponding sound pressure value. Furthermore, the feature extraction and fusion methods of welding electrical parameters and welding sound signals were investigated based on the two abnormal welding states: gas feeding error and welding wear. In the aspect of electric signal, welding current, welding voltage, line energy were studied, and in the aspect of sound signal, MFCC is extracted after de-framing and windowing. BiLSTM-CTC algorithm has been used to identify welding process gas feeding error and welding wear. For the recognition model the classification error rate based on sound feature is the lowest at 0.389, and the classification error rate based on electrical signal and acoustic signal fusion feature is at 0.295.