Weld evaluation processes are usually conducted in the post-weld stage. In this way, defects are found after the weld is completed, often resulting in disposal of expensive material or lengthy repair processes. Simultaneously, weld quality inspections tend to be performed manually by a human, even for an automated weld. Therefore, a proper real-time weld quality monitoring method associated with a decision-making strategy is needed to increase the productivity and automaticity in weld. In this study, acoustic emission (AE) as a real-time monitoring method is introduced for gas metal arc weld. The AE system is designed to cover a wide range of frequencies from 5 to 400 kHz. Additionally, the welding parameters (weld current, voltage, gas flow rate, and heat input) are recorded concurrently with AE. Different types of weld defects are artificially created to generate different signals. For the automated decision-making system, machine learning algorithms are used. Several features extracted from the AE and welding parameters feed into a machine learning algorithm. A new AE feature as the rate of AE energy accumulation extracted from time driven AE feature is defined. For decision-making, supervised learning models are trained and evaluated using testing data. General classification methods—such as Logistic Regression—predict each data-point separately. In this study, Adversarial Sequence Tagging method is applied to predict the presence of four weld states as good, excessive penetration, burn-through, porosity and porosity-excessive penetration. We explore the prediction task as a sequence tagging problem where the label of a data-point depends on its corresponding features as well as neighboring labels. When all the AE features as well as heat input are used in the feature set, the sequence tagging and logistic regression algorithms achieve a prediction accuracy of 91.18% and 82.35%, respectively, as compared to metallographic analysis.

焊接评估过程通常在焊后阶段进行。以这种方式，在焊接完成之后发现缺陷，通常导致处置昂贵的材料或冗长的修复过程。同时，焊接质量检查往往是由人工手动进行的，甚至是自动焊接。因此，需要一种与决策策略相关的适当的实时焊接质量监控方法，以提高焊接的生产率和自动化程度。在这项研究中，引入了声发射（AE）作为气体金属电弧焊的实时监控方法。AE系统设计为覆盖5到400 kHz的广泛频率范围。另外，焊接参数（焊接电流，电压，气体流速和热量输入）与AE同时记录。人为地创建了不同类型的焊接缺陷以生成不同的信号。对于自动决策系统，使用机器学习算法。从AE和焊接参数中提取的几个特征输入到机器学习算法中。定义了一种新的AE功能，即从时间驱动AE功能中提取的AE能量累积速率。对于决策，使用测试数据对监督学习模型进行训练和评估。通用分类方法（例如逻辑回归）分别预测每个数据点。在这项研究中，采用对抗序列标记法来预测四种焊缝状态的存在，即良好，过度熔深，烧穿，气孔和气孔过量熔透。我们将预测任务作为序列标记问题进行研究，其中数据点的标签取决于其相应的特征以及相邻的标签。当在特征集中使用所有AE特征和热量输入时，与金相分析相比，序列标记和逻辑回归算法的预测精度分别达到91.18％和82.35％。