A digital twin-enhanced deep metric learning framework for few-shot combustion fault diagnosis in marine diesel engines

Combustion fault diagnosis in marine diesel engines is hindered by scarce labeled fault samples, feature drift under variable operating conditions, and cross-condition distribution shifts. This paper proposes a digital twin (DT)-enhanced deep metric learning framework for few-shot combustion fault diagnosis. A 0D/1D thermodynamic DT model of a Wärtsilä 6L20 marine diesel engine is built in MATLAB/Simulink, and three typical combustion faults are parametrically injected. Five physics-informed thermodynamic features augment raw multi-sensor signals. A 1D-CNN Siamese network with a joint triplet–center loss maps input signals into a 128-dimensional embedding space optimized for inter-class separation and intra-class compactness. A two-stage training strategy pre-trains on abundant DT-generated source-condition data and fine-tunes only fully connected layers on K target-condition samples per class. Under a cycle-level, leakage-free data splitting protocol, the proposed joint-loss method achieves 87.05% ± 5.14 pp (mean ± standard deviation over 10 runs) accuracy at K = 20, while a triplet-only ablation reaches 90.91%, indicating that the triplet term is the dominant driver of few-shot discrimination. The average accuracy remains 84.43% under combined noise and speed fluctuation interference across 25–100% MCR loads. Ablation, visualization, and GUM-compliant measurement uncertainty studies validate each module. The framework’s primary contribution lies in physics-informed feature augmentation and data-efficient two-stage training rather than a claim of state-of-the-art accuracy; it provides a simulation-based foundation for training diagnostic models when real-vessel fault data are unavailable.