Alignment Correction and Systematic Bias Feedforward for Large-Format Bipolar-Plate Stacking

To meet the accuracy and takt requirements of large proton-exchange-membrane fuel-cell (PEMFC) stacks, we develop a lightweight vacuum gripper and an online, vision-based alignment scheme dedicated to large-format bipolar plates (BPPs). Using ORB plus RANSAC, in-plane offsets (Δx,Δy) are estimated and compensated in process, closing the loop of “top-view alignment → in-process pose correction → bottom-view verification.” Single-layer stacking experiments on a STEP SR20 robot show converged errors with suppressed heavy tails: medians approach zero, interquartile ranges (IQRs) shrink markedly, and the 90% containment radius decreases significantly, demonstrating robustness under heterogeneous fields of view and image sharpness. Building on this, we propose an offline “mean removal + covariance whitening” calibration and feedforward strategy to construct an isotropic, unified coordinate frame and couple it with the online loop into a low-complexity workflow without adding sensors. This provides a baseline for deviation control and process stabilization in multilayer stacking.