An unsupervised simulation-to-real framework for bearing fault diagnosis via robust GSR and lightweight domain adaptation network

Intelligent mechanical fault diagnosis has witnessed considerable progress with the application of unsupervised domain adaptation (UDA). These approaches primarily mitigate real-world cross-domain shifts, as adequate labelled fault samples are rarely available and costly to annotate in practice. Simulation generates low-cost labelled data, yet its applicability suffers from simulated-to-real domain shift induced by interference or variations in operating conditions. This study proposes an unsupervised simulated-to-real framework, SCFE-LDAN. Specifically, the signal conditioning and feature enhancement (SCFE) — built on the derived stability conditions of damping-modulated generalized stochastic resonance (DGSR) system — integrates noise injection, normalization, and DGSR-based robust feature enhancement. It unifies the processing of simulated and real data without relying on label information, effectively mitigating initial domain discrepancies while enhancing feature discriminability. Additionally, a lightweight domain adaptation network (LDAN) combines improved Softmax (I-Softmax) classification loss with a hybrid domain distribution matching (DDM) loss that fuses correlation alignment (CORAL) and maximum mean discrepancy (MMD). By integrating SCFE with LDAN, a unified two-stage UDA framework is constructed, enabling efficient knowledge transfer from labelled simulated data to unlabelled real data while maintaining model compactness. Experimental results validate SCFE-LDAN’s superiority in diagnostic accuracy, noise robustness, and adaptability to compound faults, confirming its potential value in industrial scenarios where data annotation is scarce or infeasible. • SCFE-LDAN addresses simulated-to-real domain shift, overcoming UDA limitations. • Stability-constrained SCFE unifies signal processing and enhances fault discriminability. • Lightweight LDAN fuses I-Softmax and DDM loss for efficient knowledge transfer. • SCFE-LDAN exhibits superiority for data-scarce and complex industrial scenarios.