304不锈钢快频脉冲TIG焊的熔透识别与闭环控制

Penetration Recognition and Close-loop Control Method of Fast-frequency Pulsed TIG Welding Process for 304 Stainless Steel

焊缝熔透的识别与控制对提高快频脉冲TIG焊(Fast-frequency Pulsed Tungsten Inert Gas welding,FFP-TIG welding)的焊接质量和自动化水平具有重要意义。熔池图像可以作为卷积神经网络(Convolution Neural Network,CNN)的输入来识别熔透状态,但模型往往较为复杂,难以部署于工业边缘设备中,同时缺乏基于识别结果的熔透状态控制方法。针对该问题,搭建视觉传感系统来构建304不锈钢FFP-TIG焊熔池图像数据集,包括未熔透、适度熔透和过度熔透三种标签;开发基于知识蒸馏(Knowledge Distillation,KD)的熔透识别模型,利用MobileNetV2作为教师模型,用于训练自定义的学生模型ResNet-KD;基于ResNet-KD的softmax层输出向量设计模糊控制器,实时调整快频基值电流,实现熔透状态的闭环控制。结果表明,ResNet-KD在验证集上的准确率达到了97.60%,在预设电流、变散热及热量积累的工况下,模糊控制器均有良好的性能。

Weld penetration recognition and control is of great significance to improve welding quality and automation level of Fast-frequency Pulsed Tungsten Inert Gas welding(FFP-TIG welding).Weld pool images can be used as the input of convolutional neural network(CNN)to recognize the penetration state,but the model is often complex and difficult to deploy in industrial edge devices.At the same time,it lacks control methods for penetration state based on recognition results.To solve this problem,a visual sensing system was established to construct the FFP-TIG welding pool image dataset of 304 stainless steel,including lack of penetration,desirable penetration,and excessive penetration.A penetration recognition model based on knowledge distillation(KD)was developed,which utilized MobileNetV2 as the teacher model for training ResNet-KD that is a custom student model.Based on the softmax layer output of ResNet-KD,a fuzzy controller was designed to adjust the fast-frequency base current in real-time and realize closed-loop control of the penetration state.The results showed that the recognition accuracy of ResNet-KD on the validation set achieved 97.60%,and the fuzzy controller had good performance under the conditions of presetting initial cparameters,heat dissipation variation and heat accumulation.