Recurrence plots are widely used to represent and characterize recurrence behaviors of complex systems. Although the recurrence plot captures a great deal of useful information, very little has been done to leverage convolutional neural networks (CNNs) regarding the learning of recurrence dynamics. In this paper, we develop a new CNN approach to investigate recurrence patterns in multisensor signals for real-time anomaly detection. The proposed methodology is evaluated and validated in both simulation studies and a real-world case study for quality control of a flash welding process, which is commonly used to manufacture anchor chains in the shipbuilding industry. First, we develop a new sensing system to collect electrical-current and electrode-position profiles in the flash welding process. Second, recurrence plots are derived with multisensor signals collected from the manufacturing process of each workpiece in the anchor chain. Third, CNNs are developed to learn workpiece-to-workpiece variations in the recurrence plots. Experimental results show that the proposed CNN models of recurrence plots yield superior performance for anomaly detection of welding quality variations, improving the automation level of flash welding processes.