Structural Condition Monitoring Using Deep Learning on a Metallic Part Fabricated by Additive Manufacturing

Abstract

Additive manufacturing (AM) was originally developed to manufacture polymer prototypes. Today, it has been used for the manufacturing of many critical machine components. Most of the structural health monitoring (SHM) methods were developed for monitoring the condition of large and thin plates on airplane fuselages. Additively manufactured parts are generally small, thick, and have complex geometries. SHM methods have been improved to sense load, detect defects, and identify loose bolts with the help of a permanently installed sensor. In this study, the adaptability of SHM methods was researched with additively manufactured metal parts with complex geometry. Magnets were used to apply pressure to 9 different locations on the surface of a stainless steel additively manufactured thick plate with deep groves. SHM was used to estimate the magnets’ location. Many SHM (Lamb wave) methods cannot work on smaller parts since their dimensions are shorter or very close to the wavelength of the created oscillations. Surface response to excitation (SuRE) method which has similar characteristics to electromechanical impedance methods was used for data collection. To obtain descriptive features of the time domain data, fast Fourier transformation (FFT), short-time Fourier transformation (STFT), continuous wavelet transformation (CWT), and synchrosqueezing transform (SST) were applied to CWT. 1D and 2D convolutional neural networks (CNN) were used to classify the cases. When CNN was optimized for the analysis of our data, 100% location estimation accuracy was obtained by using 50% of 320 scalograms for training. The scalograms were obtained by enhancing the CWT results with SST. STFT-CNN combination was the second best. It obtained 95% accuracy with the same number of spectrograms and training allocation.