Abstract
This paper describes a framework for using software agents to facilitate concurrent engineering of electromechanical products. By enabling agents—both autonomous programs and human interfaces—to share rich, machine-readable physical models, the system aims to accelerate the design process and reduce information latency. We introduce the Compositional Modeling Language (CML) and demonstrate its application in the redesign of an optical pick-up head (PUH) for DVD players. The goal is to move beyond simple parameter exchange to sharing the underlying logic and constraints of a design, allowing engineering teams to identify cross-disciplinary conflicts early and significantly reduce expensive late-stage backtracking.
Introduction: The Problem of Information Latency
Designing modern electromechanical products requires tight integration between mechanical, electrical, and control systems. Traditionally, specialists in these fields work in silos, exchanging information through manual reports or static data files. This creates a major bottleneck known as information latency. For example, a mechanical engineer might change a component's mass without the control engineer realizing the impact on system stability until weeks later during detailed simulation. This lack of "model-level" communication leads to a cycle of "design-test-fail-backtrack" that inflates development time and costs. The challenge is to create a system where agents and experts share not just numbers, but the assumptions and physical laws behind them.
Methodology: Agent-Based Concurrent Engineering
The researchers developed an agent-based infrastructure centered on three key components:
- Compositional Modeling Language (CML): A declarative language used to represent physical knowledge. Unlike standard simulation scripts, CML represents "model fragments" (e.g., spring physics or motor dynamics) that can be automatically composed into a system model for specific scenarios.
- Java Agent Template (JAT): A framework for creating software agents that communicate over the Internet using KQML (Knowledge Query and Manipulation Language) and KIF (Knowledge Interchange Format).
- Integrated Toolchain: The system wraps commercial engineering tools—specifically MATLAB for control design and I-DEAS for CAD—into agents. A central "CML Agent" maintains the master logical model of the product, acting as a single source of truth.
Knowledge Representation Levels
A core contribution of the paper is the categorization of design information into five levels of abstraction, ranging from high-level conceptual needs to detailed 3D geometry.
The authors argue that while the industry is proficient at sharing the "bottom" levels (parameters and geometry), the "middle" levels—mathematical equations and ontologies—are where the most valuable design agility is found. To bridge these levels, the system uses a translation agent to map local tool terminologies to a shared global ontology.
Case Study: The DVD Optical Pick-up Head
The framework was validated using the redesign of an optical pick-up head (PUH). This case study included physical parameters for 6-DOF rigid body dynamics, heat transfer models, and magnetic field non-linearities. The CML agent demonstrated its utility by providing parameters (like damping coefficients) that were calculated dynamically based on temperature dependencies defined in the physical model, rather than being static values.
Results: Performance Sizing vs. Conventional Design
The authors advocate for "Performance Sizing," a phase between conceptual and detailed design where "rough" but formally defined models are used to evaluate trade-offs. This allows for direct optimization across disciplines rather than slow iteration.
As illustrated in the figure above, the proposed approach replaces long waiting times and major "No-Go" failures with smaller, faster iterations. This significantly reduces the total design time by identifying cross-disciplinary conflicts early in the process, leading to a more efficient path to an "OK" design state.
Limitations and Future Work
While the current implementation successfully shares parameters (Level D) with shared context, the authors acknowledge that without shared assumptions (like units or coordinate systems), parameter sharing remains prone to error. Future work aims to support sharing at the equation level (Level C), allowing agents to exchange raw mathematical relationships for more powerful global optimization. Additionally, the team plans to extend the system to handle hybrid modeling, combining continuous physical dynamics with discrete control logic.
Summary updated