Detection of Anomalies in Additively Manufactured Metal Parts Using CNN and LSTM Networks

Abstract

The process of metal additive manufacturing (AM) involves creating strong, complex components by using fine metal powders. Extensive use of AM methods is expected in near future for the production of small and medium-sized batches of end-use products and tools. The ability to detect loads and defects would enable AM components to be used in critical applications and improve their value. In this study, the Surface Response to Excitation (SuRE) method was used to investigate wave propagation characteristics and load detection on AM metallic specimens. With completely solid infills and the same geometry, three stainless steel test bars are produced: one conventionally and two additively. To investigate the effect of infills, four bars with the same geometries are 3D printed with triangular and gyroid infills with either 0.5 mm or 1 mm skin thickness. Two piezoelectric disks are attached to each end of the test specimens to excite the parts with guided waves from one end and monitor the dynamic response to excitation at the other end. The response to excitation was recorded when bars were in a relaxed condition and when compressive loads were applied at five levels in the middle of them. For converting time-domain signals into 2D time-frequency images, the Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT) were implemented. To distinguish the data based on fabrication characteristics and level of loading, two deep learning models (Long Short-term Memory algorithm (LSTM) and Convolutional Neural Networks (2D CNN)) were utilized. Time-frequency images were used to train 2D CNN, while raw signal data was used to train LSTM. It was found that both LSTM and 2D CNN could estimate solid parts' loading level with an accuracy of more than 90%. In parts with infills, CNN outperformed LSTM for the classification of over five classes (internal geometry and loading level simultaneously).