Unsupervised learning for detection of defects during machining of additively manufactured components

Abstract

Detection of defects such as porosities is important in additive manufacturing to ensure the desired material properties. Most additively manufactured components require finish machining to achieve the desired surface finish and geometric tolerances. This paper presents an unsupervised machine learning approach using an autoencoder to detect defects during the machining of additively manufactured components. To train the autoencoder, machining tests at different process parameters were completed on fully dense wrought Grade 91 steel. The spindle power and the tool position were monitored during these machining tests. The autoencoder reconstruction error in the wrought machining tests was used to establish a threshold to detect defects during the machining of Grade 91 samples manufactured using directed energy deposition. Results show that the autoencoder was able to detect multiple porosities along the length of the cut at different machining process parameters with accuracy greater than 0.8. Although the machining process is destructive, it is a fast and low-cost approach for defect detection in additive manufacturing. The method can also be implemented in a hybrid setup to ensure defect-free components or for rapid identification of optimal process parameters.