Research on Fault Diagnosis Algorithm Based on Multi-mode Control System

In this paper, we propose a compact hybrid architecture that fuses a Densely Connected Convolutional Network (DenseNet) with a Transformer encoder to achieve robust fault diagnosis under low signal-to-noise ratios (0–15 dB). DenseNet’s dense connectivity enhances gradient flow and preserves fine-grained diagnostic features, while the Transformer’s multi-head self-attention adaptively highlights critical spectral components in noise-corrupted signals. We validate our method on a simulated ITC97 four-op-amp dual-second-order filter dataset generated via PSpice Monte Carlo analysis, comparing to DBN, GMKLSVM+PSO, and DSDAE baselines. Experimental results show that our DenseNet– Transformer model attains up to 98.20% accuracy. These findings demonstrate that the synergistic integration of dense connections and self-attention is key to enhancing noise resilience in fault-diagnostic applications.