Smart Diagnostics: How Argonne Could Use Generative AI to Empower Nuclear Plant Operators
30 Julio 2024 - 10:40AM
Business Wire
Imagine being able not only to detect a fault in a complex
system but also to receive a clear, understandable explanation of
its cause. Just like having a seasoned expert by your side. This is
the promise of combining a large language model (LLM) such as GPT-4
with advanced diagnostic tools.
In a new paper, engineers at the U.S. Department of Energy’s
(DOE) Argonne National Laboratory explore how this novel idea could
enhance operators' interaction with diagnostic information in
complex systems like nuclear power plants. The goal is to improve
decision-making by presenting diagnostic information in clear,
understandable terms that detail what is wrong, why it is wrong,
and how it can be addressed.
Argonne engineers combined three elements: an Argonne diagnostic
tool called PRO-AID, a symbolic engine and an LLM to achieve this.
The diagnostic tool uses facility data and physics-based models to
identify faults. The symbolic engine acts as an intermediary
between PRO-AID and the LLM. It creates a structured representation
of the fault reasoning process and constrains the output space for
the LLM, which acts to eliminate hallucinations. Then, the LLM
explains these faults in a straightforward manner for the
operators.
“The system has the potential to enhance the training of our
nuclear workforce and streamline operations and maintenance tasks,”
says Rick Vilim, manager of the Plant Analysis and Control and
Sensors department at Argonne.
PRO-AID works by comparing real-time data from the plant to
expected normal behaviors. When there’s a mismatch, it indicates a
fault. This process involves using models that simulate the plant’s
components and how they should normally behave. If something
doesn’t match, there’s a problem, and PRO-AID provides a
probabilistic distribution of faults based on these mismatches.
A key challenge with LLMs is ensuring they provide accurate
information. The authors address this by designing a symbolic
engine to manage the information the LLM uses, ensuring it only
provides explanations based on the data and models.
The LLM is used to explain the results from PRO-AID. It takes
complex technical data and translates it into easy-to-understand
language. This helps operators understand the cause of the fault
and the reasoning behind the diagnosis. Additionally, using natural
language, the operators can use the LLM to inquire arbitrarily
about the system and sensor measurements.
The research was funded by DOE’s Office of Nuclear Energy.
View source
version on businesswire.com: https://www.businesswire.com/news/home/20240730156379/en/
Christopher J. Kramer Head of Media Relations Argonne National
Laboratory Office: 630.252.5580 Email: media@anl.gov