Spatiotemporal Generative Adversarial Network for Industrial Fault Diagnosis With Imbalanced Data

High-quality data is crucial for fault diagnosis, yet industrial settings often face challenges such as data imbalance and a scarcity of labeled instances, which hinder diagnostic accuracy. Generative models have emerged as a key solution to these issues. To maximize the utility of generated data in model training, it needs to align with the original data in terms of distribution, temporal dynamics, and spatial relationships. To generate industrial time-series data, we develop a spatiotemporal generative adversarial network (GAN) that integrates three key innovations: a variational autoencoder long short-term memory generator to capture temporal dependencies, a mode-seeking regularization (MSR) technique to mitigate mode collapse, and a dual-adversarial mechanism that enhances data quality by incorporating a Siamese convolutional neural network alongside the discriminator. Recognizing that existing similarity metrics primarily focus on data distribution, often leading to a narrow evaluation scope, we propose a comprehensive similarity evaluation framework that assesses generated data across distribution, temporal, and spatial dimensions. Our method surpasses the second-best approach in all three aspects, achieving superior F1 score and G-means, thereby ensuring high-quality synthetic data for industrial fault diagnosis.