A Cycle-Consistent Generative Model for Bidirectional Cross-Modal Translation between Industrial Tabular and Time-Series Data in an Ultraviolet Lamp Pinch-Sealing Process

Abstract

Quality inspection in the ultraviolet (UV)-lamp pinch-sealing process is challenging because defects can arise from subtle deviations in process conditions and transient equipment responses during sealing. Therefore, modern production lines collect tabular process variables together with time-series sensor signals, which enables multimodal learning for defect detection and predictive maintenance. However, in real production lines, missing modalities frequently occur due to sensor failures, logging errors, or misalignment between systems, which severely affect the applicability of multimodal learning. Existing cross-modal generative models mainly target semantically aligned pairs such as image-text or image-image pairs, and fail to capture the sequential structure and information asymmetry between tabular and time-series data in manufacturing. Therefore, this work proposes a practical cycle-consistent generative framework tailored to manufacturing environments that reconstructs missing tabular or time-series modalities and enables robust multimodal learning. First, a cycle-consistent generative adversarial network (CycleGAN)-based bidirectional translation model is designed between tabular data and time-series latent representations extracted by a convolutional autoencoder (CAE), enabling training with both paired and unpaired samples. Second, a set-guided mixture-of-experts (MoE) generator is introduced on the tabular side, clustering process conditions, and assigns specialized experts to improve the fidelity of generated tabular variables. Third, a process-aware discriminator is proposed to incorporate inter-modal correlations and process labels during training, thereby enabling the generator to produce samples consistent with underlying manufacturing conditions even when labels are unavailable at inference. Fourth, a cycle-consistent distillation (CyDi) module regularizes the tabular-to–time-series mapping using intermediate features from the time-series-to-tabular mapping, mitigating the information imbalance between modalities and enhancing time-series generation quality. The effectiveness of the proposed framework is validated on a real manufacturing dataset by training downstream defect classifiers on imputed multimodal data. Experimental results show that, when using the data generated by the proposed method, various multimodal learning models achieve average performance gains of 9.15 percentage points (pp) in the area under the precision–recall curve (AUPRC) and 9.83 pp in the F1 score. The proposed method can be flexibly applied even in settings with incomplete data and limited paired samples, and it is designed to remain applicable when labels are unavailable and some modalities are missing, indicating strong practical potential for deployment in real industrial environments.