LSTM-based visual sequence regression for inner-wall roughness prediction in small-diameter CFRP holes

Abstract

Carbon-fiber-reinforced polymer (CFRP) aerostructures are susceptible to drilling-induced inner-wall roughness, which can degrade joint strength and fatigue reliability. Conventional contact profilometers are limited by low efficiency and potential surface damage, while the optical approach is difficult to apply to the confined, curved inner-wall of small-diameter holes. To address these challenges, this paper proposes a Long Short-Term Memory (LSTM)-based visual sequence regression method for the non-contact prediction of the arithmetic mean roughness (Ra) of CFRP drilled holes. Inner-wall images are acquired from four circumferential orientations, and a 1000-pixel grayscale sequence is extracted from the central region of each image as the sequential input. An LSTM regression network is then trained to learn the mapping between grayscale texture sequences and surface roughness. Experimental results on the test set demonstrate that the proposed method achieves an RMSE of 0.37 µm, an MAE of 0.30 µm, and an R2 of 0.93 for Ra prediction, indicating high accuracy and robust generalization. These results confirm the feasibility of non-contact quantitative roughness assessment using visual sequential features and LSTM regression, providing an effective pathway for in-process quality inspection of small-diameter CFRP holes.