Abstract
This paper explores an early and influential attempt to bridge the gap between design and manufacturing using Artificial Intelligence. It introduces REX (Riveting Expert), a knowledge-based system designed to automate the intricate process of engineering riveted joints in aircraft-grade aluminum.
Introduction
In traditional engineering, a significant "wall" often exists between the design phase, focused on function and geometry, and the manufacturing phase, focused on tools and shop-floor execution. While numerical tools like Finite Element Analysis (FEA) had automated stress calculations by the mid-1980s, they lacked the "experiential knowledge" or "rules of thumb" senior engineers use to ensure a part can be built efficiently. The REX system addresses this challenge of Design for Manufacturability (DFM) specifically for riveted joints, where choosing the right rivet type, spacing, and sheet overlap requires specialized knowledge to satisfy both structural integrity and factory-floor limitations.
Methodology
The authors propose a theoretical framework based on two primary hypotheses regarding expert problem-solving:
- Modular Organization: Experts solve problems through top-down abstraction (moving from general suitability to specific rivet placement) and horizontal dissection (breaking a complex joint into simpler pairwise lap joints).
- Modular Identity: Different tasks require distinct reasoning styles. REX utilizes three specific module types to handle these tasks.
REX was implemented in VAX Common LISP using a frame-based representation to store knowledge about materials (aluminum alloys), requirements (axial shear loads), and manufacturing resources (available rivet inventory and tool access).
Reasoning Modules
To mimic human expertise, REX integrates three distinct reasoning styles:
- Algorithmic: Used for numerical computations, such as calculating load-carrying capacity.
- Rule-based Selection: Used for choosing the "best" option from a set of valid candidates, such as selecting a specific joint type.
- Constraint-based Restriction: Used for pruning a search space, such as eliminating rivets that are incompatible with material thickness.
Key Insights and Results
A notable feature of REX is the "Maybe" response in constraint modules. Unlike simple binary logic, these modules can return a "maybe" accompanied by a subspace for further investigation, allowing the system to handle borderline manufacturability cases.
The system's knowledge base was built from approximately 350 hours of knowledge engineering with domain experts. In testing, REX demonstrated high efficiency and accuracy:
- Efficiency: It can produce a complete design and manufacturing specification in roughly one to two minutes of run time.
- Accuracy: In a complex test case involving a 45,000 lb load, REX correctly designed a single splice butt joint with 88 rivets in seven rows, matching expert recommendations for rivet type and spacing.
- Validation: Solutions were validated by an independent expert, proving the captured expertise was generalizable.
Conclusion
REX demonstrates that AI's value in engineering lies not just in automating single tasks, but in managing the information flow between design and manufacturing. While initially limited to flat aluminum joints, it serves as a foundational model for "intelligent automated assistants" capable of handling broader manufacturing processes like welding and bonding across various materials.
Summary updated