C-GAN-VAE: Causal Generative Adversarial Variational Autoencoder for few shot fine grained cross domain fault diagnosis for planetary gearbox

Maintaining operational reliability of industrial systems requires advanced fault diagnostic techniques. While deep learning has brought significant advances to this field, current methodologies face substantial difficulties in few-shot learning scenarios and in generalizing to previously unencountered fault types. These limitations stem from two primary factors: the substantial requirement for labeled training data that is often unavailable in practical applications, and the failure to identify and represent the fundamental causal mechanisms governing fault patterns. To address these critical challenges, this paper introduces the Causal Generative Adversarial Variational Autoencoder (C-GAN-VAE) framework. The key innovation is a unified loss function combining reconstruction, adversarial, code prediction, and causal disentanglement objectives in single-stage training, unlike existing approaches that either neglect causality or employ complex multi-stage procedures. Our architecture incorporates a specialized Causal Q Network (CQN) that disentangles latent representations into independent factors corresponding to distinct fault-related causal mechanisms, substantially improving novel fault recognition capabilities. The framework further integrates a conditional Generative Adversarial Network (CGAN) that produces high-quality synthetic samples to augment limited fault data, thereby strengthening model generalization under data scarcity. Comprehensive evaluation performed on the Spectra Quest Inc. (SQI) wind turbine drive-train simulation test rig dataset and the Case Western Reserve University (CWRU) dataset validates the exceptional performance of C-GAN-VAE in cross-domain and cross-machine fault diagnosis under few-shot constraints. The proposed approach demonstrates consistent performance superiority, achieving accuracy improvements of 2-5% in 5-way classification tasks and 3-7% in more complex 10-way tasks across both 1-shot and 5-shot learning configurations when measured against current benchmarks.