Digital twin-driven few-shot fault data generation and intelligent diagnosis for rolling bearings

This is the authors' abstract. We don't add key points for this paper.

In the context of limited data availability under few-shot learning conditions, we propose an innovative fault data generation and intelligent diagnosis method for rolling bearings driven by a Digital Twin (DT) framework. To address limited fault samples, a Convolution-Bi-directional Long Short-Term Memory (Conv-BiLSTM) network integrates virtual and real-world signals to generate simulated vibration signals. Time–frequency characteristics are leveraged to reconstruct diverse fault scenarios for inner rings, outer rings, and rolling elements from a single fault signal, enriching the dataset. a Secretary Bird Optimisation Algorithm (SBOA) optimises a Kolmogorov-Arnold Network (KAN) for intelligent diagnosis using the generated data. Experiments show our method outperforms traditional approaches, achieving 97.9% accuracy. Validated on two datasets, it demonstrates robustness and practicality for few-shot fault diagnosis.