Start Date

19-5-2023 3:00 PM

End Date

19-5-2023 3:15 PM

Document Type

Short Paper

Keywords

Structural Health Monitoring, Additive Manufacturing, Convolutional Neural Network, Long Short-Term Memory

Description

An active Structural Health Monitoring (SHM) method called Surface Response to Excitation (SuRE), is used in this study to detect and quantify the damages created by a milling operation on additively manufactured metal plates. This method entails bonding two piezoelectric disks to the test specimens, one to excite it with surface waves from one end of the plate, and the other to sense the dynamic response to excitation at the other end. A sweep sine wave with a duration of 1 ms, ranging from 50-120 kHz is used as the excitation signal. Five stainless steel plates of identical size (195×54×2.5 mm) were created using a Markforged metal 3D printer. The data for four different conditions were recorded, which are, when the parts were undamaged and when they were face milled at 3 different lengths. The data was then used to train One-Dimensional and Two-Dimensional Convolutional Neural Networks (CNN) and also a Long Short-Term Memory (LSTM) neural network. The continuous wavelet transform (CWT) was used to convert the collected sensor data from the time domain to time-frequency representation images to be classified by the 2D CNN. The 1D CNN, 2D CNN and the LSTM classified the damage length with an overall accuracy of 98.9%, 100%, and 97.8% respectfully.

DOI

https://doi.org/10.5038/ZPYX4570

Share

COinS
 
May 19th, 3:00 PM May 19th, 3:15 PM

Damage identification in 3D printed metal parts using deep learning

An active Structural Health Monitoring (SHM) method called Surface Response to Excitation (SuRE), is used in this study to detect and quantify the damages created by a milling operation on additively manufactured metal plates. This method entails bonding two piezoelectric disks to the test specimens, one to excite it with surface waves from one end of the plate, and the other to sense the dynamic response to excitation at the other end. A sweep sine wave with a duration of 1 ms, ranging from 50-120 kHz is used as the excitation signal. Five stainless steel plates of identical size (195×54×2.5 mm) were created using a Markforged metal 3D printer. The data for four different conditions were recorded, which are, when the parts were undamaged and when they were face milled at 3 different lengths. The data was then used to train One-Dimensional and Two-Dimensional Convolutional Neural Networks (CNN) and also a Long Short-Term Memory (LSTM) neural network. The continuous wavelet transform (CWT) was used to convert the collected sensor data from the time domain to time-frequency representation images to be classified by the 2D CNN. The 1D CNN, 2D CNN and the LSTM classified the damage length with an overall accuracy of 98.9%, 100%, and 97.8% respectfully.