Quality Monitoring in AM Metal Through Efficient Streaming Machine Learning

Abstract

Additive Manufacturing (AM), popularly known as 3D printing, has become a transformative technology with wide-ranging applications in a number of industries such as medical, aerospace, and automotive. The increasing adoption of AM in these critical sectors necessitates a focus on ensuring the quality of printed components. An emerging method to estimate the quality of an object is to process data coming out of the printer during printing. This data can then help detect defects within the object. Most of the current research focuses on the ability and accuracy of finding defects but fails to take into account the crucial time aspect associated with processing the data coming from the printer. This thesis proposes a method of quality monitoring in AM through the integration of machine learning and low-latency datastream processing. Monitoring live sensor data from inside the printer (in-situ) has been advocated to minimize time for defect detection, yet its successful implementation relies on establishing a quantitative relationship between sensing data and the defects in an object. This proves challenging given the multitude of variables involved with this specific type of AM method.

This thesis utilizes Optical Tomography (OT) images captured as the object is being printed. The OT images are used together with spatial algorithms and machine learning models to identify defects in objects. The entire process is implemented in a streaming fashion, enabling low-latency monitoring and assessment. Importantly, the proposed approach supports not only the detection and position of defects but also the size of the defects, addressing both aspects that are critical for assessing the quality of AM components. The proposed methodology contributes to the advancement of in-situ quality control in AM, addressing the critical need for timely defect detection and ensuring the production of high-quality components. The output includes the location of defects, and the geometric attributes of porous areas across the current layer, considering the adjacent layers in the object during the manufacturing process.

The proposed method for finding defects has a sub-quadratic processing overheadin the size of the input. In this work it was tested with data of multiple objectsand images for 250 layers; with a recall of up to 95%, its running time for finding overlapping clusters in two adjacent layers is a few milliseconds, hence allowing significant margin for closed-loop control.