Session: 02-02-02: Session #2: Measurement Science, Sensors, Non-destructive Evaluation (NDE) and Process Control for Advanced Manufacturing

Paper Number: 96010

96010 - Heterogeneous Sensing and Bayesian Optimization for Smart Calibration in Additive Manufacturing Process 

Fused Filament Fabrication (FFF) is an extrusion-based additive manufacturing process that utilizes a filament material melted through a hot end extruder to generate a component. Despite the great potential of the process to drastically reduce time-to-produce, cost and material waste for the creation of geometrically complex components, the presence of diverse defects deteriorate the quality of the final build. Defects in FFF (e.g., voids, stringing, and varying track width) are primarily linked to improper calibration of parameters, including feed speed, extrusion speed, extruder temperature, and build plate temperature.

Trial and error is the most common practice implemented to manually offset baseline parameters using an array of components generated with varying process parameters. However, fabrication with manual adjustment not only is time consuming, but also leads to a suboptimal solution that jeopardizes the strength and integrity of the generated components. Furthermore, this adjustment faces transferability issues across design and material (i.e., new calibration is required to change materials or design).

We propose a novel Bayesian Optimization (BO) methodology in conjunction with heterogeneous sensing to determine optimal process parameters with a minimum number of experiments. BO consists of two steps: First, a Gaussian Process as a surrogate model that maps the relationship between controllable parameters (e.g., feed rate/flow rate ratio, extrusion temperature, and height offset) and build quality (i.e., the objective function that is derived from sensing data). Second, an acquisition function is defined from this surrogate to decide where to sample.

Specifically, the Gaussian process provides a Bayesian posterior probability distribution that describes potential values for objective function at a candidate set of process parameters. Each time we observe an objective function at a new set of values for process parameters, this posterior distribution is updated. The acquisition function measures the value that would be generated by evaluation of the objective function at a new set of values for process parameters, based on the current posterior distribution over the objective function. Here, the build quality is formulated as an objective-scoring algorithm that returns the proportion of the effective specimen sensor measurements divided by the desired values. This multi-fidelity scoring algorithm incorporates a regularizer vector that maps the effectiveness of each type of sensor in capturing specific types of defects and is learned through maximum posteriori estimation or maximum likelihood method.

Before a print is begun, a calibration operation is completed where a small test component is created. During the creation of this component, filament extrusion speed, feed rate, extruder temperature, and height offset is adjusted according to the acquisition function and surrogate model until the quality score is raised to the desired amount. The derived optimal process parameters and surrogate model are then used in the following component print. During the print, the quality score is observed. If the quality score falls below the desired threshold, the print is paused, and a calibration operation is again completed to return to optimal process parameters. This methodology is implemented on the open-source software of OctoPrint, allowing for easy replication for researchers and hobbyist printers.

The experimental results are conducted on a Prusa 3D printer integrated with a vibration sensor (i.e., HiLetgo GY-291 ADXL345 3-Axis Digital Acceleration) and two types of cameras (i.e., Adafruit MLX90640 IR Thermal Camera and NoIR Camera Module V2) in addition to preexisting extruder temperature sensors. The results show that optimal process parameters for novel filament materials can be determined in a fraction of the time compared to manual calibration, drastically reducing the number of trials required to achieve optimal build quality. The platform created to implement this optimization solution has the potential to not only greatly improve novel filament material testing, but also enable the widespread use of proposed optimization models in other additive manufacturing processes such as laser powder bed fusion and direct energy deposition.

Presenting Author: Farhad Imani University of Connecticut

Presenting Author Biography: Farhad Imani joined the University of Connecticut in 2020 as an Assistant Professor in the Department of Mechanical Engineering. He received his Dual-title Ph.D. in Industrial Engineering and Operations Research from the Pennsylvania State University in 2020. His research interests focus on data analytics, machine learning, statistical learning, and decision theory for process monitoring and control, system diagnostics and prognostics, quality and reliability improvement with applications in advanced manufacturing.

Authors:

Sean Rescsanski University of Connecticut
Mahdi Imani Northeastern University
Farhad Imani University of Connecticut