Manufacturing Quality Inspection of Electric Vehicle Reducer Gears Using CUDA-Parallel Order-Domain Feature Extraction

Electric vehicle reducer gears directly affect drivetrain noise, vibration, and harshness; furthermore, manufacturing-related anomalies must be inspected before their assembly. This study proposes a CUDA-parallel order-domain feature-extraction framework for the manufacturing quality inspection of electric-vehicle reducer gears under variable-speed dynamic meshing conditions. Speed-pulse signals are used to reconstruct the shaft angle, and vibration signals are transformed into the order domain through computed order tracking. Sixteen physically interpretable features are extracted from the gear meshing orders, ghost orders, order-band energies, and angle-domain statistics. The proposed method is evaluated using 640 DMTD samples featuring five gear states. File-level stratified cross-validation indicates macro-F1 scores of 0.9936 and 0.9886 for acceleration and deceleration conditions, respectively. The CUDA implementation yielded speedups of 45.9× for angular resampling, 49.7× for feature extraction, and 5.40× for overall computation (excluding I/O).