Conditional generative adversarial network-enhanced multi-layer parallel pooling vision transformer for federated prediction of product quality

In the decentralized manufacturing context, the production management system uses multiple sensors to record Multi-dimensional Time-varying Series Data (MTSD) from machines geographically distributed. These data share the same sample space but have different feature spaces, characterized by scarcity, imbalance, and noise. To address the common challenges of limited overlapping data between participants and class imbalance in the decentralized manufacturing context, this study proposes a Vertical Federated Learning (VFL) framework for predicting product quality, together with a Conditional Generative Adversarial Network(cGAN)-enhanced Multi-Layer Parallel Pooling Vision Transformer (MLP-PiT) model. The VFL approach eliminates the need for independent external coordinators, with one data holder (active party) coordinating training while others (passive parties) participate, aligning with the constraints of decentralized manufacturing environments. The cGAN-based data augmentation technique uses conditional vectors to produce additional overlapping data for training models across participants. This reduces overfitting caused by limited data in the VFL setting and improves training efficiency for the MLP-PiT model. Comparative experiments and data analysis are conducted with experimental data collected from distributed product assembly. The results confirm the feasibility and effectiveness of the proposed method.