Industrial chemical facilities produce vast amounts of data, yet fault detection remains challenging due to dynamic processes and non-linear interactions. This study proposes a novel self-supervised learning (SSL) architecture using prediction contrast learning to enhance fault detection & diagnosis(FDD) in industrial settings. Our model integrates time and frequency domain information to capture crucial temporal dependencies and improve robustness. We perform complex network topology analysis for cloud and edge computing environments, considering industrial networks as complex systems with non-trivial topological features. Experimental results on the Tennessee Eastman Process Ricker dataset show significant improvements, achieving a True Positive Rate of 92% and reducing Average Detection Delay to 22.18 s. This approach addresses data scarcity issues and enhances fault detection accuracy in complex environments.