Wire arc additive manufacturing (WAAM) has received attention because of its high deposition rate, low cost, and high material utilization. However, quality issues are critical in WAAM because it builds upon arc welding technology, which can result in low precision and poor quality of the melted parts. Hence, anomaly detection is essential for identifying abnormal behaviors and process instability during WAAM to reduce the time and cost of post-process treatment. The relevant studies have been conducted on anomaly detection algorithms using machine learning in fused deposition modeling and laser powder bed fusion; however, they have less investigated the implementation for in situ quality monitoring in WAAM. This work presents a real-time anomaly detection method that uses a convolutional neural network (CNN) in WAAM. The proposed method enables creation of CNN-based models that detect abnormalities by learning from the melt pool image data, which are pre-processed to increase learning performance. A prototype system was implemented to classify melt pool images into “normal” and “abnormal” states, with the latter accounting for balling and bead-cut defects. Experiments were conducted using molybdenum, a cost-intensive and hard-to-machine material. Four CNN-based models were created using MobileNetV2, DenseNet169, Resnet50V2, and InceptionResNetV2. Then, their performances were validated in terms of classification accuracy and processing time. The MobileNetV2 model yielded the best performance with 98 % of classification accuracy and 0.033 s/frame of processing time. This model was also compared with an object detection algorithm named “YOLO”, which yielded 73.5 % of classification accuracy and 0.067 s/frame of processing time.

电弧增材制造（WAAM）因其高沉积率、低成本和高材料利用率而受到关注。然而，质量问题在 WAAM 中至关重要，因为它建立在电弧焊技术之上，这可能导致熔化零件的精度低和质量差。因此，异常检测对于识别 WAAM 期间的异常行为和过程不稳定性以减少后期处理的时间和成本至关重要。在熔融沉积建模和激光粉末床融合中使用机器学习的异常检测算法进行了相关研究；然而，他们对 WAAM 中原位质量监测的实施研究较少。这项工作提出了一种在 WAAM 中使用卷积神经网络 (CNN) 的实时异常检测方法。所提出的方法能够创建基于 CNN 的模型，这些模型通过从熔池图像数据中学习来检测异常，这些图像数据经过预处理以提高学习性能。实施了一个原型系统，将熔池图像分类为“正常”和“异常”状态，后者解释了成球和切珠缺陷。使用钼进行了实验，钼是一种成本密集且难以加工的材料。使用 MobileNetV2、DenseNet169、Resnet50V2 和 InceptionResNetV2 创建了四个基于 CNN 的模型。然后，在分类准确性和处理时间方面验证了它们的性能。MobileNetV2 模型以 98% 的分类准确率和 0.033 秒/帧的处理时间产生了最佳性能。该模型还与名为“YOLO”的目标检测算法进行了比较，得到了 73 个结果。