Abstract
Reducing the setup time for a new production line is critical to the success of a manufacturer within the current competitive and cost-conscious market. To this end, being able to reuse already available machines, toolings and parts is paramount. However, matching a large warehouse of previously engineered parts to a new component to produce, is often more a matter of art and personal expertise rather than predictable science. In order to ease this process we developed a database retrieval approach for mechanical components that is able to deal with both rigid matching and deformable shapes. The intended use for the system is to match parts acquired with a 3D scanning system to a large database of components and to supply a list of results sorted according with a metric that expresses a structural distance.
Introduction
In large-scale manufacturing, such as eyewear production, the transition to a new product line involves significant costs and time delays. A major factor in these overheads is the creation of new "toolings"—the custom jigs, molds, and machine configurations required for specific parts. Reusing existing toolings from previous production cycles can drastically reduce these costs.
Historically, identifying a "similar enough" part within a warehouse containing tens of thousands of components has relied on "tribal knowledge." Factories often depend on veteran engineers to recall similar designs from memory. This approach is inherently unscalable and vulnerable to the loss of expertise through retirement. This paper presents an automated 3D shape-retrieval pipeline designed to transform this process from an intuitive "art" into a predictable, data-driven science.
The Retrieval Pipeline
The authors propose a multi-stage pipeline that transitions from physical parts to a searchable digital database. The process begins with 3D acquisition and ends with a ranked list of similar components based on structural distance.
3D Acquisition
To populate the database, parts are scanned using a custom structured-light system integrated with an automated turntable. This setup allows for a full 360-degree reconstruction of a component in approximately two minutes, ensuring high-fidelity digital twins of physical inventory.
Semi-supervised Segmentation
A significant challenge in mechanical part retrieval is the accurate segmentation of complex assemblies (e.g., separating a hinge or a nose pad from a glasses frame). Fully automated methods often struggle with the intricate geometries of mechanical parts. To address this, the authors developed a "human-in-the-loop" tool.
The system utilizes a dual graph representation of the mesh. A user provides initial "seeds" or hints, and the algorithm performs greedy region growing. Edge weights in the graph are determined by surface curvature, causing the algorithm to "slow down" at sharp edges or notches—natural boundaries between functional components.
Matching Strategies
The pipeline employs two distinct matching techniques depending on the nature of the component:
- Rigid Matching: For unarticulated parts like hinges, the system uses a variant of the Iterative Closest Point (ICP) algorithm. To prevent the algorithm from converging on local minima, Principal Component Analysis (PCA) is used to provide a robust initial alignment.
- Non-rigid Matching: For deformable parts like earpieces, the authors introduce a novel approach inspired by bioinformatics. The 3D component is sliced along its median axis into a sequence of 2D cross-sections. These slices are treated as a "genetic sequence" of geometric signatures. The system then applies the Needleman-Wunsch and Smith-Waterman algorithms—standard tools for DNA alignment—to find the best match between shapes, even if they are bent or deformed differently.
Evaluation and Results
The system was evaluated using an industrial dataset from Luxottica, comprising approximately 30,000 eyeglass models and nearly 150,000 individual components.
- Efficiency: The system can search the entire database of 150,000 parts in approximately 10 seconds on a standard server. This speed is achieved by an initial filtering step that removes parts whose total surface area deviates by more than 30% from the query.
- Accuracy: For rigid components, the "correct" match was typically the top-ranked result. For non-rigid components, the Smith-Waterman local alignment successfully placed a relevant match within the top 50 results in over 90% of cases.
- Industrial Validation: Expert feedback during a multi-month trial confirmed that the system consistently surfaced relevant parts capable of reducing tooling costs.
Conclusion and Future Work
The proposed pipeline successfully bridges the gap between physical inventory and digital retrieval. By leveraging bioinformatics techniques for non-rigid geometry, the authors provide a scalable solution for manufacturing reuse. Future improvements may include automating the initial segmentation seeding and incorporating manufacturing constraints—such as material properties or CNC machining compatibility—into the similarity metric to further enhance practical utility for engineers.
Summary updated