TLDR: A multi-agent AI system has been developed for CERN’s UNICOS industrial control system to address key operational challenges. This AI assistant automates the decoding of complex widgets, performs root cause analysis by traversing device hierarchies and leveraging live data, and traces datapoint elements responsible for widget animations. The system significantly reduces cognitive load for operators, speeds up diagnostic processes, and improves maintainer efficiency, demonstrating the practical application of AI, multi-modal reasoning, and Retrieval Augmented Generation (RAG) in critical industrial control environments.
CERN, the European Organization for Nuclear Research, is at the forefront of scientific discovery, and its operations rely heavily on sophisticated control systems. One such system is the UNified Industrial Control System, or UNICOS, which manages over 200 critical systems, including the technical infrastructure for the accelerator complex. While powerful, UNICOS has presented several challenges for its operators and maintainers, leading to increased cognitive load and slower diagnostic times in critical situations.
A recent project has introduced an innovative AI-enhanced operator assistant designed to tackle these pain points. This assistant aims to streamline operations, reduce manual effort, and improve overall efficiency within the UNICOS environment. The core of this solution is a multi-agent AI system, built with a modular architecture that integrates seamlessly with existing UNICOS components.
Addressing Key Challenges with AI
The project specifically targets three major areas where UNICOS operators and maintainers faced difficulties:
1. Widget Decoding: UNICOS panels display numerous widgets, each with symbols that operators must frequently reference documentation to understand. This manual decoding process can be cognitively demanding, especially during critical alarms. The AI assistant automates this, reducing the mental burden on operators.
2. Root Cause Analysis: Diagnosing the source of an issue in UNICOS is a time-consuming task. Devices are represented at a very atomic level, creating complex hierarchies that operators must manually traverse to trace problems. The AI assistant leverages live device data and documentation to automate root cause analysis, speeding up diagnostics.
3. Tracing Datapoint Elements (DPEs): Maintainers often struggle to identify which DPEs are responsible for animating specific parts of a widget, a process complicated by a complex codebase. The AI assistant automates the tracing of these DPEs, significantly reducing debugging time.
How the AI Assistant Works
The system comprises several key components:
- A UNICOS-side extension, written in CTRL code, which includes a new AI Assistant user interface and REST API endpoints to access real-time widget information.
- A Python-based multi-agent system deployed on a virtual machine, housing a supervisor agent and three specialized worker agents: an AI Widget Decoding Agent, an AI Root Cause Analysis Agent, and an AI Widget Part Animation DPE Tracing Agent.
- A vector database that stores operator documentation and widget animation code, enabling semantic searches for relevant information.
The AI Widget Decoding Agent uses advanced image processing techniques, including image segmentation with OpenCV2 and image upscaling with an Enhanced Deep Residual Networks (EDSR) model, to accurately extract data from widget images. It then uses a powerful language model (llama-4-maverick-17b-128e-instruct) and few-shot prompting to interpret the widget’s meaning by querying the documentation stored in the vector database.
For root cause analysis, the AI Root Cause Analysis Agent interacts with newly created UNICOS REST API endpoints to retrieve live data about widget hierarchies, frontend status, and device status. It then consults the documentation to interpret these codes and identify potential issues, offering sensible suggestions for further investigation.
The AI Tracing of Widget Part Animation DPEs Agent navigates a complex codebase, also stored in the vector database, to pinpoint the exact DPEs responsible for widget animations. It uses semantic search and method-name-based queries to understand the code logic and explain how each DPE affects the widget’s appearance.
Also Read:
- FAIR-RAG: Enhancing LLM Accuracy with Evidence-Driven Iterative Refinement
- Simplifying Network Operations: How AI is Answering Questions About Network Behavior
Impact and Future Potential
Preliminary evaluations have shown that this AI-enhanced system successfully decodes widgets, performs accurate root cause analysis, and traces DPEs across the codebase. These capabilities significantly reduce the manual workload for operators and maintainers, enhance situational awareness, and accelerate responses to alarms and anomalies.
Beyond these immediate benefits, this project highlights the immense potential of integrating multi-modal reasoning (combining text and image understanding) and Retrieval Augmented Generation (RAG) into industrial control systems. It provides a robust foundation for developing more intelligent operator interfaces at CERN and paves the way for broader applications of assistive AI in accelerator operations.
This work is more than just a proof of concept; it’s a practical step towards advancing intelligent operator interfaces at CERN, offering a modular, extensible, and practically integrated AI solution. For more detailed information, you can refer to the full research paper.
