Ultrasonic phased array measurement & compression for in-process weld bevel estimation

• Real-time FMC pipeline validated for both conventional V-groove and 120 mm thick mock narrow gap geometries. • 88% FMC data compression with minimal SNR and accuracy loss. • Achieves ± 0.4° angular precision and < 2 mm bevel length error for conventional V-groove. • GPU implementation delivers sub-250 ms execution per scan frame for conventional V-groove. • Demonstrates deployable ultrasonic pipeline for automated weld inspection. This work presents a real-time ultrasonic inspection pipeline for weld bevel orientation measurement in robotic welding applications, addressing two major barriers to industrial deployment: reliable bevel geometry estimation in challenging environments and efficient management of high-throughput ultrasonic data. While phased array ultrasonics provides volumetric imaging where optical methods fail, its application in robotics has been limited by data scale and latency. Validated across both conventional V-groove and 120 mm thick narrow-gap geometries, the integrated GPU-accelerated orientation estimator and lightweight AUC compression strategy deliver ± 0.4° angular precision for conventional V-grooves; the pipeline achieves sub-250 ms execution for single-probe V-groove scanning and sub-1500 ms for complex dual-tandem narrow-gap configurations, maintaining over 88% data reduction with geometric errors below 3 mm. The system is validated on a robotically deployed phased array platform dynamically scanning stepped bevel geometries, simulating realistic industrial scenarios. Rather than focusing on novel algorithms in isolation, the contribution lies in demonstrating, for the first time, a modular and interpretable FMC pipeline that achieves real-time weld geometry inference in tandem with data sparsification. This proof-of-concept highlights a viable pathway toward embedded, closed-loop ultrasonic inspection for robotic welding automation.