Intelligent Detection of Liquid-Solid Interface Quality of Ultra-Clean Pumps for Semiconductor Manufacturing

In the field of semiconductor manufacturing, ultra-clean pumps are widely used to supply high-purity and highly corrosive liquids. The quality of the liquid-solid interface affects the working condition and service life of the pump. Due to improper processing of liquid-solid interfaces, the polytetrafluoroethylene (PTFE) bellows pumps, one type of ultra-clean pump, are prone to crack or cause solvent leaking during long-term use. In this paper, the short-wave infrared light (SWIR) with a wavelength between 1250-1350nm was employed to effectively and non-constructively penetrate ultra-clean PTFE materials and detect the inner liquid-solid interface quality of bellows. Several common defects of the liquid-solid interface were summarized, and a liquid-solid interface defect detector was trained based on the latest YOLOv8 model. In addition, to address the issue of insufficient sample size often encountered when applying deep learning in subdivided industrial fields, unsupervised learning techniques were employed to design a Generative Adversarial Network (GAN) to generate more bellows samples, which can effectively improve mean Average Precision (mAP) in this work.