Penetration/keyhole status prediction and model visualization based on deep learning algorithm in plasma arc welding
Penetration/keyhole status prediction and model visualization based on deep learning algorithm in plasma arc welding
复制标题
DOI：
10.1007/s00170-021-07903-9
复制
发表时间：
2021-08
期刊：
影响因子：
--
通讯作者：
C. Jia;Xinfeng Liu;Guodong Zhang;Yong Zhang;Chang-Hai Yu;Chuansong Wu
中科院分区：
文献类型：
--
作者：
C. Jia;Xinfeng Liu;Guodong Zhang;Yong Zhang;Chang-Hai Yu;Chuansong Wu
小孔状态的准确预测是实现小孔等离子弧焊接过程闭环控制、高效获得全熔透焊接接头的关键。从顶部视觉捕获的熔池图像提供了液态金属以及小孔行为的足够信息。熔池、等离子弧、小孔入口等特征在小孔加工过程中有明显的区别。提出了基于深度学习算法自动提取图像特征，而不是手动选择特征参数。由于使用所获取的数据直接训练深度CNN（卷积神经网络）模型会导致收敛失败，因此采用了一个经过良好训练的广义模型，并进行了相应的微调，以更容易地提取K-PAW图像特征。使用获得的数据集进行模型训练，该数据集以熔池图像作为输入，熔深/小孔状态（具有盲孔的部分熔深或具有贯穿小孔的完全熔深）作为输出。建立了熔深/小孔状态和上部熔池图像之间的潜在相关性。为了进一步验证训练模型的有效性和可靠性，设计了实验，获得了典型的恒定焊接电流下的小孔慢转和脉冲焊接电流下的小孔快转。基于给定的数据，验证的90%的准确度实现正确预测的小孔/渗透状态。最后，对卷积层进行了可视化，清晰地展示了卷积层的特征，这对于理解神经网络的内在机理具有重要意义。
Accurate keyhole status prediction is critical for realizing the closed-loop control of the keyhole plasma arc welding (K-PAW) processes for acquiring full-penetration weld joints with high efficiency. Visually captured weld pool images from topside provide sufficient information of the liquid metal as well as keyhole behaviors. Weld pool, plasma arc, and keyhole entrance could be clearly recognized reflecting the different features during different keyholing stages. It was proposed to extract the image features automatically based on a deep learning algorithm rather than manually selecting characteristic parameters. Since directly training the deep CNN (convolutional neural network) model using the acquired data led to convergence failure, a well-trained generalized model was employed and fine-tuned accordingly to more easily extract the K-PAW image features. Model training was conducted using obtained dataset, which took weld pool images as input and penetration/keyhole status (partial penetration with a blind keyhole or full penetration with a through keyhole) as output. Underlying correlations between the penetration/keyhole status and topside weld pool images were established. For further verifying the effectiveness and reliability of the trained model, experiments were designed acquiring typical slow keyholing under constant welding current and rapid keyhole switching under pulse welding current. Based on the given data, the verified 90% accuracy was achieved for correctly predicting the keyhole/penetration status. Finally, the visualization of the convolutional layers was carried out, and displayed the features clearly, which is of great significance for understanding the internal mechanism of the neural network.