Dry-Type Transformers Fault Early Warning Based on SSTN-DBSCAN

To address the limitations of traditional fault diagnosis methods for urban rail transit rectifier transformers—including reliance on single-source signals, weak anti-interference capability, and poor adaptability of fixed thresholds—this study proposes an early fault warning method based on a Self-Supervised Training Network (SSTN) and Adaptive Density-Based Spatial Clustering (DBSCAN). Utilizing vibration signals as supervisory inputs, a self-supervised autoencoder model extracts mechanical state-related features from acoustic signals to achieve signal enhancement. The method integrates unsupervised clustering algorithms with Principal Component Analysis (PCA) for dimensionality reduction and adaptive clustering analysis, while incorporating a dynamic threshold optimization mechanism to improve anti-interference alarm performance. Experimental results demonstrate a fault detection accuracy of 96.2%, a false-alarm rate below 2%, and significantly enhanced robustness in complex noise environments. This approach provides efficient and reliable technical support for intelligent operation and maintenance of urban rail transit rectifier transformers, offering substantial engineering application value.