Enhanced Quality Assurance via In-situ monitoring of Selective Laser Melting

Quality assurance and process reliability remain a huge challenge in SLM machines. In a study by Sirris machine learning algorithms were used to treat the melt pool signatures and layer control systems data to improve both aspects by means of melt pool monitoring and the layer control system.

Despite many industrial technological advancements in selective laser melting (SLM) machines, the quality assurance and reliability of the process is a major challenge. To overcome this challenge, the commercial SLM systems are equipped with in-situ sensing modules that effectively monitor the process at various levels such as powder-bed spreading, melt pool monitoring and laser power monitoring. But the lack of data treatment and absence of link among monitored process signatures hinders the exploitation of the full capability of these systems. As the origin of the defect in the final manufactured part is not linked to independent process signature rather inter-dependent on other process signatures as well.

To illustrate this, let us assume there is a bad powder spreading in a specific layer due to part hopping. This phenomenon will strongly influence the melt-pool signature as well. Therefore, monitoring the independent process signature will not be able to predict the final quality of the printed part. Therefore, the linking between process signatures is an inevitable need of the time. In our study, we make use of the machine learning algorithm to treat the melt pool signatures and layer control systems data to improve the process reliability and quality assurance.

Melt Pool Monitoring

The Melt Pool Monitoring (MPM) suite installed on in-house available commercial SLM 280HL was used to capture the MPM signatures and the sensitivity analysis of the equipped system was performed. As the enormous size of the data captured during the process possesses a major challenge.  To overcome this challenge, we applied the Support Vector Machine (SVM) algorithm on the extracted features from the MPM dataset. SVM was trained on the balanced dataset and tested for artificially mimicked defects. For example, the figure below shows the location of the lack of fusion defect in the part.

(a) Predicted layers for the lack of fusion defect, (b) optical micrograph of the defect

Layer Control System

Similarly, we used the layer control system installed on SLM 280HL to monitor the powder bed spreading. The images captured after every pass of the powder spread were processed using Convolutional Neural Network (CNN) to detect the various powder spreading failures such as part hopping, recoater damage, non-uniform spreading, .etc. With the help of the CNN algorithm, we successfully demonstrate the failure of the overhang part as shown in the next figure a (failure of part after layer number 300). In the figure below we also, demonstrate the link between the powder-bed spread and the MPM process signature to predict the failure in the overhang part.

(a) Powder bed spread failure predicted by CNN, (b) Melt Pool failure predicted by SVM