​The focus of our project was to conduct a feasibility study into applying Explainable AI (XAI) to hydrogen production as part of a Decision Support System (DSS). The DSS would help operators with the day-to-day running of the production facility, providing recommendations which would result in a more optimised facility. In particular, the project aimed to provide benefit for the control room operator – the person(s) managing the system alarms.

​

Consortium details

​

Intelligent Plant

Lead organisation, specialising in real-time industrial data and providers of the Industrial App Store.

​

European Marine Energy Centre (EMEC)

Provided hydrogen expertise and data from real-world assets, based in Orkney.

​

University of Aberdeen (UoA)

Contributed AI expertise to apply real-world heuristics of hydrogen production for AI learning.

 

Key Project Deliverables

​

• Construction of a domain which can be run via a simulation.

• Gather rules from domain experts to understand how they make decisions and encode these in a machine interpretable format.

• Creating the underlying formal framework where machine reasoning will take place and create tools to interrogate the system and understand these reasoning processes.

• Development and completion of a DSS dashboard.

 

Firstly, what is XAI?

​

Explainable AI (XAI) models provide detailed explanations for their answers. Trained on specific domains, they allow users to understand how conclusions are reached.

​

XAI builds trust by explaining its reasoning, unlike traditional AI models that give answers without explanations. This understanding is crucial for trust.

​

Hydrogen production involves complex systems with many variables. XAI can help operators make better decisions quickly, improving efficiency and safety.

​

Why focus on alarms?

​

Hydrogen’s unique properties and storage conditions pose significant risks. Explainable AI could enhance safety by addressing these risks, especially given the rapid development and deployment of new systems.

​

The project focused on alarm management and control room operator guidance due to the consortium’s expertise in this area, as well as the broader applicability of these solutions.

​

Hydrogen systems use various industrial control systems, from small units in devices like electrolysers and compressors to large, complex networks in production plants.

​

Control room operators manage complex displays with constantly changing data and many controls. During issues like equipment failures or shutdowns, alarms can flood the screens with flashing warnings and messages about sensor problems. This can cause stress on the operator, and can increase the chance of human error coming into play.

​

This issue affects not only the hydrogen sector but also other industries with complex control systems.

Given the risks of handling hydrogen, safe operation is crucial. Using AI can help by simplifying the information presented to operators, improving decision-making, quality, safety, and reducing downtime. Explainable AI allows operators to understand and challenge AI decisions, providing additional benefits.

​

What did we work on?

​

Firstly, we created a simple training environment for both humans and AI to operate, simulating a basic hydrogen production system. The goal for the AI was to keep the system running without shutting down. The team, experienced in alarm systems and hydrogen production, identified the necessary components for the environment.

In the simulator, it is necessary to prevent tanks from overfilling, as failure to do so would stop the production. The AI’s score increased when tank levels were within safe limits, and a higher score indicated better performance. Alarms were triggered when tanks exceeded limits. The AI model had visibility on valve and alarm states, as well as tank levels.

​

The process data (valve openings, tank levels, etc) and Alarm & Event data (alarms when the tank levels exceeded normal limits and any actions that the AI took) were obtained through the AI model training and input into the Industrial App Store for use on key analytical and visualisation applications.

​

Making use of LLMs

​

With the data available to examine through these applications, LLMs were introduced into the project. These were used to explain the alarm and event data in SEER – an application that visualises how past events are connected. During alarms, users need quick explanations. LLMs translate SEER’s data into easy-to-read text, helping users understand and respond faster.

​

It is recommended that control systems have an Alarm Response Manual (ARM), detailing every alarm within system, explaining actions and consequences. If uploaded to an LLM, it could provide quick, clear responses during alarm situations, potentially saving lives. Combining an ARM with SEER data would offer deeper insights and early guidance.​