Penn State is building factory AI that explains its advice and learns engineers’ rules
Two manufacturing case studies will test an approach that links learned patterns with engineering knowledge. The work remains under development, not a released system or a proven performance gain.
The project pairs neural pattern recognition with knowledge graphs and engineer feedback, targeting recommendations that can be checked and revised rather than acted on automatically. Its tests span weld-defect diagnosis at a Pennsylvania manufacturer and task rebalancing for General Motors; both are research case studies, not operational deployments. Penn State says the work is funded through the Department of Energy’s Genesis Mission, with federated learning and before-and-after energy measurements also under study—important tests of whether the method can transfer across shops without exposing raw production records.
01
For welding, the proposed system could trace detected porosity to possible causes such as electrode travel speed or voltage, then offer a setting for a person to assess.
02
In GM’s planning case, rejecting a task move because a tool is only available at one station could teach the system a previously unstated constraint.
03
Federated learning would let shops train locally and share model learnings instead of raw production records, addressing data-sharing concerns and limited training data.
Penn State researchers are building factory AI meant to explain why it recommends a change—and learn when an engineer says it will not work. In a November 16 research update, the university detailed testing in weld inspection and General Motors assembly-line planning. The work remains under development, not a released product or completed deployment.
Led by industrial engineering professor Soundar Kumara, the project receives funding through the U.S. Department of Energy’s Genesis Mission. Penn State first described its selection for Phase I funding in July. The latest update lays out two industrial case studies and how the team intends to connect machine learning with engineers’ practical knowledge.
Connecting a detected problem to a reason
The team calls its approach neurosymbolic intelligence: neural networks learn patterns from industrial data, while a separate map holds engineering facts and rules. That map, known as a knowledge graph, links items such as defects, machine settings and production steps. The intended result is a recommendation whose reasoning an engineer can inspect.
Engineers would be able to add rules as they learn them. Operators would also be able to review what the neural system flags and override it with help from the engineering knowledge built into the system. Rather than asking people to accept a detected pattern alone, the design aims to give them a basis for checking it.
At a Pennsylvania manufacturer, students are training neural models to recognize weld defects, including cracks, porosity and undercut. They are also building knowledge graphs from welding rules and standards. The researchers say the planned system should connect a defect with its likely cause and suggest an adjustment for a person to assess and implement if appropriate.
For example, porosity detected in a weld image could lead the system through its rules to a possible problem with electrode travel speed or voltage. It would suggest a setting for the welder to check, rather than make the adjustment itself. The team also intends the system to help train newer welders as experienced workers retire.
Editorial illustration for Penn State is building factory AI that explains its advice and learns engineers’ rules.Source: psu.edu.
The physics and engineering knowledge we build into the symbolic side gives us a way to test the AI, not just train it.
Rebecca Napolitano, project co-principal investigator and Penn State associate professor of architectural engineering
Teaching a planner the factory’s unwritten rules
The General Motors case study tackles a different problem: redistributing tasks among assembly-line workstations when the product mix or required speed changes. Moving one task can force other moves because tasks depend on one another. Plants use optimization models—software that searches for the best plan within specified rules—but those rules may leave out practical constraints.
In the proposed approach, an optimization model supplies a plan and a neural layer learns from engineers’ responses. An engineer might reject a task move because the necessary tool is available at only one station. The system is intended to infer that unwritten constraint, add it to the planning model and generate a revised proposal.
Anirudh Subramanyam, who leads the optimization effort with GM’s Seog-Chan Oh, describes the goal as learning rules that exist only in engineers’ heads. The system should then explain which constraint shaped its advice. The researchers deliberately chose a small manufacturer and a large one to examine how their approach works at different scales.
Learning across shops without pooling production records
The team is also exploring federated learning, in which each shop trains on its own data and shares what its model learned rather than raw production records. Kumara says small manufacturers often lack enough data to train a good model alone and are reluctant to hand production data to larger companies or competitors.
Alongside the AI work, Penn State’s Pennsylvania Technical Assistance Program will measure shop energy use before and after the tools are introduced. That gives the project a planned physical check on what changes at the factory, beyond whether the software can produce an explanation.
Reader comments
Newest comments first. Replies stay oldest first.