Self-Supervised Contrastive Learning with Dynamic Cost-Sensitive Random Forest for Intelligent Fault Diagnosis of Diesel Engines

Reliable identification of incipient degradation is critical to the durability and functional integrity of diesel engines. However, diagnosis is challenged by scarce degradation labels, severe class imbalance, and heterogeneous multi-source monitoring data. This study proposes a selfsupervised contrastive learning with dynamic cost-sensitive random forest (SSCL-DCSRF) framework for intelligent fault diagnosis with a focus on wear- and lubrication-related degradation states. Rather than directly measuring surface morphology, roughness, residual stress, or wear loss, the proposed method identifies operating states associated with surface-related degradation from multi-source monitoring data. A physically consistent view-generation strategy preserves operating-condition semantics and cross-variable relationships, enabling robust representation learning from unlabeled data. A dynamic cost-sensitive random forest then integrates class frequency, feature-space dispersion, and prediction uncertainty to improve the recognition of minority and difficult-to-classify states. Experiments on a multi-source diesel-engine fault dataset demonstrate that SSCL-DCSRF outperforms representative machine-learning methods in class-balanced diagnostic performance, with enhanced sensitivity to early wear and lubrication-related degradation. The proposed framework provides an effective data-driven approach for indirect identification of operating states related to surface degradation and integrity-oriented maintenance of critical tribological interfaces.