Convolutional Neural Network Based Image Recognition for Crack Severity Detection and Service Life Assessment of Dissimilar Metal Welded Joints

Dissimilar aluminum-steel joining is essential for lightweight multi-material structures but is limited by a narrow process window in which lack of fusion at low heat input competes with interfacial degradation at high heat input. This study presents an integrated framework combining process-window mapping, interfacial metallurgy, ordinal crack-severity grading using convolutional neural networks (CNNs), and fatigue-based service-life assessment for laser-joined Al-steel joints. A well-defined stable processing window was identified between under-processed and crack-prone regimes, characterized by a thin and continuous intermetallic layer (3.4 ± 0.5 μm), whereas over-processing resulted in significant thickening (11.8 ± 2.1 μm) with microcrack and void-chain formation. Mechanical performance showed a strong monotonic degradation with increasing crack severity, confirming severity as a physically meaningful descriptor of structural integrity. An ordinal CNN model achieved robust grading performance (quadratic-weighted κ = 0.91, macro-F1 = 0.86, AUROC = 0.94) with reliable confidence calibration (ECE = 0.024), enabling decision-oriented inspection. Fatigue analysis further demonstrated systematic life reduction with increasing severity, with clear separation in survival behavior and hazard escalation across grades. By integrating severity prediction with confidence-gated screening, the proposed framework reduced the estimated field failure rate from 12% to 2% (83% reduction). These results establish a practical and interpretable pathway for linking process control, microstructural evolution, and AI-assisted inspection to improve manufacturability, reliability, and lifecycle performance of dissimilar metal joints.