Data-Driven Virtual Temperature Sensor for Monitoring of Vacuum Reflow Soldering Processes

Precise temperature control is essential in surface mount technology (SMT) reflow soldering, particularly in vacuum soldering processes where thermal deviations can lead to defects such as void formation, incomplete soldering, and reduced solder joint reliability. In practice, direct temperature measurement at all critical locations within the soldering chamber is often infeasible due to sensor placement constraints and process interference. This thesis presents a data-driven temperature prediction approach for non-contact thermal monitoring in a batch-type vacuum reflow soldering system. Time-series data collected from multiple thermocouples during controlled heating cycles are used to train recurrent neural network (RNN) models, including vanilla RNN, GRU, and LSTM architectures, that estimate temperatures at inaccessible locations on the soldering plate. The trained GRU model, selected as a trade-off between prediction performance and model complexity, provides real-time predictions of the local temperature evolution based solely on measurements from accessible sensors. Experimental results show that the proposed virtual sensing approach accurately captures the thermal dynamics of the reflow process at all stages of the temperature profile. By eliminating the need for physical sensors in critical locations, the method improves process observability, thermal control, and quality assurance in advanced vacuum reflow soldering applications.