Introduction
Mold manufacturing is a fundamental pillar of industrial production, responsible for creating the intricate components—such as core pins, cavities, and lifters—required for mass-production molds. Traditionally, managing the mold component machining workflow is a labor-intensive endeavor that relies heavily on the intuition and experience of human experts. These experts must manually analyze 3D CAD models to identify required machining processes (e.g., grinding, electrical discharge machining), estimate processing times, and coordinate complex machine schedules to maintain factory efficiency.
IMPsys (Intelligent Mold Processing System) addresses the lack of integrated automation in this domain. While prior research has targeted specific aspects like 3D design or quality control, IMPsys provides an end-to-end pipeline that automates the transition from raw 3D design files to optimized production schedules. Currently deployed in the South Korean manufacturing sector, the system successfully replaces manual expert labor with a streamlined, AI-driven workflow.
System Architecture and Methodology
The IMPsys architecture is designed as a comprehensive stack for hosting services, integrating database clusters and message queues to manage the flow of data between three core AI-driven modules: the Job Explorer (JE), the Process Time Inferrer (PTI), and the Job Scheduler (JS).
Job Explorer (JE)
To avoid the computational overhead of direct 3D mesh labeling, the JE module transforms 3D models into six 2D "side view" images. A YOLOv5 object detection model is then employed to identify specific machining areas and classify them into six primary types, such as Grinding, EDM, and Wire EDM. This approach leverages mature computer vision architectures to handle 3D spatial data efficiently.
Process Time Inferrer (PTI)
The PTI module utilizes a multivariate regression model to predict machine occupancy times. It extracts geometric features from 3D models, including total volume, point/cell counts, and "cut volume" (the difference between the model's volume and its bounding box). To capture visual complexity, the system uses Principal Component Analysis (PCA) to generate image embeddings, which are processed through a neural "Feature Net."
Job Scheduler (JS)
Scheduling is framed as a Job-Shop Scheduling Problem (JSSP). The authors implemented an "exhausting greediness" algorithm that assigns jobs to the latest available devices while adhering to strict machining sequences. This ensures that prerequisite steps, such as grinding before EDM, are strictly maintained in the production timeline.
Implementation and Results
The system was trained using custom datasets derived from real-world manufacturing data. The JE module was trained on 3,108 2D images from 518 3D mold files, while the PTI module utilized historical processing times recorded by factory workers.
IMPsys was deployed at UJU Electronics in the Republic of Korea. Key performance metrics include:
- Detection Accuracy: The YOLOv5-based JE module achieved near-perfect accuracy in identifying machining parts and types in test scenarios.
- Estimation Precision: The PTI module maintained a prediction error of less than 0.5 hours, providing a reliable foundation for scheduling.
- Operational Efficiency: The automated scheduling significantly reduced the manual labor required for factory floor management.
Discussion and Conclusion
A critical insight of IMPsys is the efficacy of 2D projections for 3D tasks, particularly the importance of capturing concavity in viewpoints to reflect tool accessibility. By combining geometric statistics with learned visual embeddings, the system captures both the physical bulk and the structural complexity of mold components.
While highly effective in its current deployment, future work could enhance the Job Scheduler to handle more non-standard machining requirements and explore advanced scheduling algorithms beyond the current greedy approach. Nonetheless, IMPsys represents a significant step forward in the realization of fully automated smart factories for mold manufacturing.
Summary updated