Fluid catalytic cracking (FCC) is a pivotal secondary processing unit in the petroleum refining industry, primarily designed to convert heavy hydrocarbon fractions into high-value light products via catalytic reactions. The reliable operation of FCC units is paramount for global energy supply and economic efficiency, necessitating robust and timely process safety management. However, industrial FCC units operate under severe conditions characterized by high temperature, short reaction time and significant dynamic disturbances. This non-stationary characteristic presents a severe domain adaptation challenge for conventional data-driven monitoring schemes. The intrinsic coupling between the reaction-regeneration system and downstream fractionation units introduces substantial process variability and complexity. Furthermore, the stringent safety constraints of chemical facilities often result in a scarcity of labeled fault data, complicating the development of robust intelligent monitoring systems. To address these challenges, this work proposes a novel fault detection framework, VATE, tailored for FCC processes under non-stationary operating conditions. This study first utilizes a rigorous FCC process simulator to synthesize a comprehensive fault dataset, incorporating stochastic feedstock fluctuations to mimic real-world scenarios. In the proposed architecture, a variational mechanism is integrated into the attention module to achieve Variational Attention. The VATE framework leverages the probabilistic modeling capability of the variational mechanism to handle data uncertainty, while the attention mechanism focuses on relevant fault-indicative features. This structure enhances the model's ability to capture latent process features robust to fluctuations, thereby improving fault detection performance. A transformer encoder serves as the backbone for data reconstruction, using its intrinsic capability to model long-range temporal dependencies in the multivariate time-series data, while kernel density estimation (KDE) is employed to establish the statistical threshold strategy for anomaly detection. The VATE method was validated on a constructed catalytic cracking simulation dataset. When compared to the performance of principal component analysis (PCA), autoencoder (AE), variational autoencoder (VAE), long short-term memory autoencoder (LSTM-AE), long short-term memory variational autoencoder (LSTM-VAE), and dot-product attention transformer encoder (DATE), VATE achieved the best fault detection performance of 94.64% across eight fault types, providing valuable insights for future intelligent process monitoring of catalytic cracking processes.