穿孔等离子弧焊接(K-PAW)能够在不开坡口时一次焊透10mm厚度以上的钢板，在中厚板构件焊接加工领域具有重要应用潜力，但小孔贯穿熔池的复杂行为规律导致该工艺合理参数范围较窄、焊缝熔透不稳定。本项目提出一种基于CNN+LSTM深度学习提取熔池图像时空关联特征的新方法，实现对穿孔等离子弧焊接穿孔/熔透状态的准确预测。基于CNN卷积神经网络对正面熔池图像的静态空间特征进行自动提取，并将序列特征信息输入到LSTM长短期记忆神经网络，提取历史动态图像的时间特征；建立CNN+LSTM深度学习数学模型和框架，通过建立多参数、多类型数据集对模型进行训练、调优，实现同时基于历史记忆和当前熔池图像特征对背面熔透状态的准确预测；开展深度学习模型可视化研究，探索图像时空关联特征在焊接过程中的实际物理意义。研究结果对于实现自动化和智能化的等离子弧焊接工艺、推动中厚板焊接生产效率的大幅提升具有重要意义。

Keyhole plasma arc welding (K-PAW) can penetrate steel plates above 10 mm thickness without any groove, which makes this advanced technology has great potential in welding production of middle-thickness structures. However, the complicated keyhole behaviors through the plate cause narrow reasonable parameter ranges and less predictable penetration. In order to realize accurate keyhole/penetration status prediction in K-PAW, it is proposed in this project that a combined deep learning algorithms CNN+LSTM can be developed to extract the spatial-temporal features of the captured images of topside weld pool. CNN (Convolutional neural network) deep learning algorithm is used to automatically extract the spatial features of the input images. LSTM-NN (Long short-term memory neural network) is proposed to obtain the temporal features of the sequential historical input data from the CNN. Multi-parameter and multi-type experiments will be conducted to establish a database with input and output image pairs. The model is trained, tested and verified using the obtained data to predict the penetration. Visualization of the deep learning model is conducted to investigate the physical meanings of the extracted spatial-temporal features in real welding processes. The research is of great significance for realizing automatic and intelligent plasma arc welding and promoting considerable advancements in welding production efficiency of middle-thickness plates.