Pulsed Infrared Thermography with Unsupervised Machine Learning for Imaging of Microscopic Subsurface Defects in Metals

We have developed a Pulsed Infrared Thermography (PIT) system with Spatial Temporal Denoised Thermal Source Separation (STDTSS) unsupervised machine learning (ML) algorithm for processing of thermal images. The PIT imaging technique is based on recording material surface temperature transients with infrared (IR) camera following thermal pulse delivered on material surface with a white light flash light. Information in surface temperature transients can reveal presence of internal defects, such as voids with lower thermal conductivity than the host material, because thermal resistance of defects results in relatively slower local decay of material surface temperature. PIT imaging has a number of potential advantages in detection of subsurface pores in metals produced with laser powder bed fusion additive manufacturing method. However, limited imaging resolution of existing systems must be increased substantially. The challenge of detection of microscopic material defects in PIT images involves indentifying weak thermal signals with intensity comparable to IR camera noise level. STDTSS algorithm developed for processing of thermograms involves wavelet transform preprocessing, followed by principal component and analysis (PCA) and independent comonent analysis (ICA). We show that STDTSS recovers microscopic defects which are not visible in recorded PIT images. In this study, stainless steel 316 (SS316) and Inconel 718 (IN718) specimens were developed with a pattern of subsurface calibrated flat bottom hole (FBH) defects with diameters created with EDM (electron discharge machininig) drill. Selecting appropriate IR camera magnification lenses and imaging frame rate, and processing PIT data with STDTSS algorithm, we show that defects as small as 200µm are dtected in SS316 and IN718 specimens.