Thesis Format

Integrated Article

On-line monitoring, neural networks-based modelling, and reinforcement learning control of laser remelting process

Author

Honghe Wu, Western UniversityFollow

Degree

Master of Engineering Science

Program

Mechanical and Materials Engineering

Supervisor

Tutunea-Fatan, O.Remus

2nd Supervisor

Bordatchev, Evgueni

Affiliation

National Research Council of Canada

Co-Supervisor

Abstract

Laser polishing (LP) is a novel advanced surface finishing process aimed to address major shortcomings of traditional polishing processes used in the tool and die industry. In LP, a concentrated energy source (laser) is used to create a localized melt pool allowing material redistribution under surface tension and various thermodynamic phenomena. The faster speed, higher areal locating precision, and the lack of consumable tooling for surface polishing by laser remelting (SP-LRM) process makes this technology highly attractive. The primary factor limiting wide-range industrial use of the LRM is complex process parameter optimization and process instabilities. In this work, reinforcement learning is applied to achieve in-situ monitoring and control of the melt pool stability using co-axially mounted high-speed near-infrared (NIR) camera and dynamically varying the laser power. The controller autonomously achieved a 20% reduction of standard deviation value of melt pool intensity over time and a 70% reduction in resultant surface roughness (Sa) variations.

Summary for Lay Audience

In the manufacturing industry, a smooth surface is often required to improve product ergonomics, appearance, and reduce friction on mechanical assemblies. This creates a need to smoothen or polish a 2-dimensional surface as fast and as consistent as possible. Existing polishing process such as abrasive sanding is slow and inconsistent due to the abrasives wearing down over time. Laser polishing (LP) is a process where a laser beam is used to melt the top surface of the material to allow the molten liquid to relocation, naturally forming a smoother surface. The process can be understood similarly to how the hockey rink is made. Hot water is used to melt the uneven ice and allowed to cool down to eventually freeze. The hot water fills in gaps and trenches in the uneven ice surface under surface tension, forming a smooth ice surface in the end. LP relies on the same concept to allow molten material (plastic or metallic) to relocate. LP is not only much faster than other conventional polishing processes, but also much more consistent while not requiring any consumables or generating any wasted byproduct. However, the melting process is highly unstable in some cases and typically requiring trial-and-error to find optimal process settings. This research applies artificial intelligence to develop an automated, self-learning controller to stabilize the melting process to ensure an ideal polished surface and stable process. The controller will use cameras to gain insight to the process condition and apply corrective control through changing the laser power during the polishing process. The final controller showed promising results, delivering a more stable process, yielding significant surface quality improvements.

Recommended Citation

Wu, Honghe, "On-line monitoring, neural networks-based modelling, and reinforcement learning control of laser remelting process" (2023). Electronic Thesis and Dissertation Repository. 9855.
https://ir.lib.uwo.ca/etd/9855

Creative Commons License

This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 4.0 License.