Spatiotemporal prediction and mechanisms of molten pool instability in variable polarity plasma arc robotic welding via CNN-LSTM

This study proposes a method for spatiotemporal prediction of molten pool states via an end-to-end CNN-LSTM model, addressing the dynamic and complex manufacturing scenarios under variable polarity plasma arc (VPPA) robotic welding. The model utilizes CNN to extract spatial features from molten pool images and employs LSTM to extract temporal features in image sequences of the molten pool. This enables early warning of transition from stability to instability of the molten pool states. Experimental results show that when predicting molten pool states at a 1.5 s prediction time using a 0.5 s image sequence sample, the CNN-LSTM model achieves a prediction accuracy of 99.21 %, with a false negative rate of only 0.72 %. In real manufacturing scenarios, the model predicts molten pool that was not part of the training data, achieving a prediction accuracy of 90.61 %. The prediction accuracy was improved to 96.43 % by fine-tuning the model with data not included in the training process. Grad-CAM visualization analysis reveals that the CNN-LSTM model primarily focuses on the rear wall region of the molten pool during the prediction of molten pool states. Insufficient molten metal supply in this region is identified as the key cause of molten pool instability. The proposed model demonstrates well performance in prediction accuracy, false negative rate, and applicability. It provides a robust method for enhancing the intelligence and reliability of VPPA robotic welding processes.