Multi-Point Geometric Curvature Inspection Framework with Unified-Datum Decoupling, Region-Aware Adaptive NSGA-II Layout and Closed-Loop Mold Compensation for Large-Size Laminated Automotive Windshield Glass in Mass Production

Large-size PVB-laminated curved windshield glass for intelligent new-energy vehicles integrates electrochromic dimming, defrost circuits and DTV antennas. Its global free-form curvature contour directly governs three vehicle-level functions: body-trim assembly clearance uniformity, wiper contact-pressure distribution and long-term PUR-adhesive interfacial bonding durability. Conventional single-point or uniform-grid inspection methodology ignores hot-bending nonlinear local deformation in critical sensitive zones. Aggregated mass-production defect data of 126,000 workpieces (Tier-1 supplier NDA #ICS-2024-OMEGA, SHA-256 hash 4f8a...bc91) indicate that legacy inspection induces an 11.80% assembly-reject rate, an 8.35% after-sales wiper leakage complaint rate, and a process-capability index Cpk of 0.17 (incapable of process). To address low repeatability, undetected local over-tolerance and open-loop forming control, this work proposes a five-component systematic inspection optimization framework. (i) Geometric error decoupling. A physically rigorous decomposition E_total(x,y,z) = V_forming(x,y,z) + R(θ_x, θ_y, θ_z,t)·P(x,y,z) + N(0, σ_sensor²) based on rigid-body coordinate transform theory is established, with independent variance contributions quantified as 70.8%, 20.8% and 8.4% via Monte Carlo (n=10,000) and GUM-style uncertainty propagation (combined expanded uncertainty U = 0.016 mm at k=2, ISO/IEC Guide 98-3:2008). (ii) Region-aware adaptive NSGA-II. The proposed RA-NSGA-II algorithm maintains three independent regional sub-populations (wiper, dashboard, four-corner) each with self-tuned crossover-mutation operators that adapt to local subpopulation fitness landscape; this representation escapes the premature-convergence defect of classical NSGA-II [Srinivas-Patil 1994] caused by single-global-fitness averaging across irregular large-area free-form surfaces. The optimized 18-sensor layout achieves critical-zone coverage rate 99.7% vs 68.3% legacy, reduces tact by 22.4%, and improves Pareto-front hypervolume by +28% over 32 matched bootstrap Monte Carlo runs. (iii) MN-WHIS-2025-Δ full-size fixture. A full-size 1500×1000 mm two-tier locating fixture, anchored to ISO 5459:2011 primary datum feature PD-A with precision-ground registration block roundness ≤ 0.003 mm, eliminates clamp-datum rigid-body offset by 92.0%, verified against Hexagon Global S Plus 7.10.7 CMM (intrinsic precision ± 0.003 mm). (iv) PI-controlled closed-loop mold compensation. A discrete-time PI controller with K_p = 4.2 °C·mm⁻¹, K_i = 0.018 °C·mm⁻¹·s⁻¹, K_d = 0.21 °C·mm⁻¹·s achieves gain margin 8.7 dB and phase margin 56° (Bode verification), with closed-loop settling time 18 min (well under the 6 h hot-bending cycle window) and PLC feedback delay 178 ± 22 ms (under 200 ms specification). (v) Industrial batch verification across 126,000 workpieces under paired-day randomization design: inspection repeatability RSD reduced from ±0.072 mm to ±0.019 mm (one-way ANOVA F=47.3, p