Open-set fault diagnosis of rolling bearings is pivotal for the simultaneous detection of known and unknown anomalies in the intelligent operation and maintenance of industrial equipment. Conventional deep-learning approaches exhibit overfitting and limited sensitivity to unknown faults in open-world environments. To overcome these limitations, this study introduces a deep energy-learning framework that integrates a dual-weight adaptive residual network with adversarial training. The proposed architecture incorporates a dual-weight-aware residual network that leverages channel attention mechanisms and hierarchical weight-modulation modules to adaptively calibrate feature responses. Adversarial training and energy-based loss functions are refined by employing projected gradient descent (PGD) attacks, parameterized by global weight coefficients, to generate adversarial samples. This mechanism enforces energy-boundary constraints, enhancing the model’s discrimination of Out-of-Distribution (OOD) instances. A novel OOD detection paradigm grounded in energy sparsity is proposed, which jointly optimizes weight-modulation parameters and feature sparsity constraints to establish dynamically calibrated energy-decision boundaries. Experimental validation on benchmark datasets demonstrates that the method achieves 99% accuracy for known faults, 93.37% AUROC for unknown faults, and 96.67% overall accuracy in mixed-sample scenarios, outperforming traditional methods. These findings provide an effective solution for industrial fault diagnosis with robust generalization and open-set identification capabilities.
An et al. (Sat,) studied this question.