The increase in complex workpieces with changing geometries demands advanced control algorithms in order to achieve stable welding regimes. Usually, many experiments are required to identify and confirm the correct welding parameters. We present a method for controlling laser power in a remote laser welding system with a convolutional neural network (CNN) via a PID controller, based on optical triangulation feedback. AISI 304 metal sheets with a cumulative thickness of 1.5 mm were used. A total accuracy of 94% was achieved for CNN models on the test datasets. The rise time of the controller to achieve full penetration was less than 1.0 s from the start of welding. The Gradient-weighted Class Activation Mapping (Grad-CAM) method was used to further understand the decision making of the model. It was determined that the CNN focuses mainly on the area of the interaction zone and can act accordingly if this interaction zone changes in size. Based on additional testing, we proposed improvements to increase overall controller performance and response time by implementing a feed-forward approach at the beginning of welding.

中文翻译：

---

卷积神经网络的远程激光焊接过程中的功率控制

随着几何形状的变化，复杂工件的增加需要先进的控制算法，以实现稳定的焊接状态。通常，需要许多实验来识别和确认正确的焊接参数。我们提出了一种基于光学三角测量反馈的，通过PID控制器通过卷积神经网络（CNN）通过卷积神经网络（CNN）控制远程激光焊接系统中激光功率的方法。使用累积厚度为1.5毫米的AISI 304金属板。测试数据集上CNN模型的总准确度达到94％。从焊接开始，控制器达到完全熔深的上升时间不到1.0 s。使用梯度加权的类激活映射（Grad-CAM）方法来进一步了解模型的决策。已经确定，CNN主要集中在交互区域的区域，并且如果此交互区域的大小发生变化，则可以相应地起作用。基于其他测试，我们提出了改进措施，通过在焊接开始时实施前馈方法来提高整体控制器性能和响应时间。