Laser material processing (LMP) has revolutionized modern manufacturing, emerging as a cornerstone technology valued for its precision, speed, and versatility, especially in high-stakes sectors like aerospace, automotive, and medical device production. Despite these advantages, techniques such as laser welding and laser powder bed fusion face substantial obstacles: ensuring consistent product quality, extensive process parameter optimization, and the need for real-time defect mitigation. This thesis systematically addresses these challenges by enhancing automation and progressively increasing the autonomy of laser processing systems.

The research begins with the development of ThreshMorph, a signal segmentation technique that improves Intersection over Union (IoU) by 39% in real-world industrial conditions compared to traditional methods. This method establishes a solid foundation for subsequent offline and online monitoring techniques.

Building on this groundwork, an interpretable adaptive filtering method for acoustic emission signals is introduced, achieving an F1-score of 97.84% in keyhole detection and an average of 85.45% in distinguishing between conduction, stable keyhole, and unstable keyhole regimes. This method directly correlates acoustic events with physical melt pool phenomena, offering transparent insights into melt pool dynamics and substantially enhancing the accuracy of regime transition detection.

To address the complexities of multi-laser systems used for higher productivity, DUAL DISCO (Dual Unmixing Acoustic Learning to Disentangle Interwoven Signal Complexity in Operations) is developed. This signal processing method effectively disentangles acoustic emission signals from multiple lasers, maintaining a regime detection accuracy of over 90% in high-throughput environments and directly tackling productivity challenges.

An uncertainty-driven iterative strategy is then proposed for autonomous parameter optimization, achieving an F1-score of 89.2% in identifying melting regimes without requiring labeled data. Through targeted experiment selection, this approach reduces experimental efforts by up to 67%, with only an 8.88% performance trade-off compared to traditional full factorial designs. This reduction in experimentation significantly improves efficiency, cuts costs, and promotes sustainability by minimizing resource consumption and waste.

Finally, an adaptive controller based on reinforcement learning is developed for real-time interaction with the laser system. Implemented on field-programmable gate array hardware for low-latency control, the controller continuously adapts by being trained by an external server without prior knowledge or extensive tuning, outperforming optimal constant-power strategies by up to 22.58% when faced with varying and irregular surface topographies. Post-process analyses confirm improved weld penetration and overall quality.

Collectively, these advancements address critical challenges in LMP by enhancing automation and introducing higher levels of autonomy. By focusing on process monitoring, parameter optimization, and real-time control, this work paves the way for more reliable, efficient, and autonomous laser manufacturing processes, meeting the demands of modern industry.