Self-attention embedded StyleGAN for virtual sample generation in sensing applications

Abstract

Given the challenges of low variability in industrial processes, which intensify data scarcity and produce anomalous distributions that compromise data-driven model accuracy. Existing sample generation methods often overlook key factors such as sparsity and correlation among data. To address these challenges, this paper proposes a StyleGAN-based virtual sample generation method with an embedded self-attention mechanism (SASG-VSG). Firstly, StyleGAN is used to map the original data space to a disentangled latent space. The output variables then act as control conditions, guiding the model to interpolate along the output dimension to ensure a more uniform distribution of generated samples. Besides, a self-attention module is incorporated into the discriminator to enhance its ability to capture the similarity between the virtual samples and the original data distribution. Finally, validation experiments on a purified terephthalic acid (PTA) solvent system and a sulfur recovery unit (SRU) confirm the capability of the proposed SASG-VSG in generating high-quality virtual samples for soft-sensing applications.