This paper introduces the VCformer, a Variational Mode Decomposition-Convolutional Neural Network-Transformer model, optimized for fault diagnosis in nuclear steam turbines amidst noisy environments. The VCformer employs VMD to break down vibration data into five distinct modal signals, and these signals, along with the original noisy input, are individually processed through a dual-layer CNN-Transformer framework, enabling precise feature extraction and fault classification for each modal component. Following this individual analysis, the VCformer utilizes an ensemble learning strategy, where a voting system dynamically integrates the outcomes from the dual-layer CNN-Transformer framework of each signal. This process effectively adjusts the voting weights in real-time, culminating in an optimized final output that leverages the comprehensive insights gained from both the original and decomposed signals. Empirical testing has demonstrated the VCformer's robust performance, with a minimal decrease in accuracy of only 0.16% even at a 20% noise level. This innovative methodology highlights the VCformer’s capability to provide accurate diagnostics in environments with significant noise, representing a novel advancement in intelligent monitoring technology for the energy sector.