Like many in industry, Chevron has experienced operators with decades of knowledge. When those operators leave, so does that knowledge and experience. With the help of Honeywell, Chevron is digitally capturing that knowledge for future operations. At the 2025 Honeywell User Group conference, Amit Jain, simulation advisor of simulation and senior process control engineer at Chevron, and Raam Thotakura, global strategic solutions leader for chemicals and sustainable energy at Honeywell, discussed their partnership and work with the Experion HALO AI-based alarm guidance to capture this knowledge. The system is not only collecting operator knowledge for the future but also making current operators more efficient by leveraging facility-wide data on alarms and outcomes in real-time.
Traditionally, process control provides a first layer of protection against anomalies in the operation. The process control system can open or close valves or speed up or slow down pumps, for example, to bring the facility back to the proper state. If not, alarms are triggered before safety is compromised.
Alarm analysis can also take up significant operator time and effort. If operators have seen the alarms in the past, they generally know how to respond quickly, but if they haven’t seen such an alarm, they need to spend time analyzing the cause. “The response the operator takes is very much dependent on their knowledge and experience,” Jain said, and the company wants to leverage that knowledge for its entire operations.
Data without context
Chevron facilities collect this data about alarms, operator responses and outcomes. “There is a lot of richness of information there. What if we can analyze all this data, five to 10 years’ worth? And from there, this is the knowledge and experience of our operators. They know the processes inside and out,” Jain said.
About eight months ago, Chevron began this process and has quickly built a library that’s capturing and applying operator knowledge across the facility. The data analysis includes not only capturing all the operator actions and outcomes, but parsing those outcomes, so it identifies the optimal response to each alarm situation.
Initially, Chevron started data analysis with pure machine learning, but all the time-based data wasn’t enough to understand the process as a whole. “In operations, it’s never one thing that’s going on. There are always multiple things going on,” Jain said. “What we saw with machine learning alone is we got a lot of hallucinations.”
Interdependent processes can make it difficult to understand what caused an alarm and how the outcomes of response actions affected the overall process. “So, if an alarm comes in and and operator an action, that action might be in response to this alarm, or it might be from something happening somewhere else in the facility,” Jain said. The machine learning models alone could not differentiate the relationships between all operators, assets and alarms across the facility.
“What are we missing? What we are missing is the context,” Jain said. To gather that context, Chevron and Honeywell worked together to build a knowledge graph for the facility, which essentially is a nodal map of the relationships between everything in the facility, including physical equipment, the control system and safety procedures (what alarms could trigger a shutdown of what processes).
One of the biggest challenges, Jain said, is building that knowledge graph. “We now have one of the most robust knowledge graphs that have been built,” he said. “It accounts for a lot of different patterns. It accounts for your asset hierarchy. It accounts for your control system. It accounts for your physical connectivity.”
Knowledge graph closes the loop
With the knowledge graph, Chevron can leverage all the information about how equipment, processes, operator actions and outcomes are impacted by each other. Then, combined with all the data about every alarm and response action, they can build a hybrid knowledge graph/machine learning approach.
“Now we have actual information from every alarm, what happened, what actually worked and what did not,” Jain said.
Armed with that historical knowledge and operational context, the models can analyze the data to determine the most optimal action for the operators. Using contextualized machine learning, Chevron built an AI-enhanced library of the optimum operator responses to specific alarms, but it did not take the models analysis at face value.
The model analysis then gets human verification, similar to the alarm rationalization process that an operator might go through, but driven more by real-time data, site-specific processes and company-wide operator knowledge.
Once a human in the loop approves the recommended action for a specific alarm situation, then it goes to the operators via the console. “It’s not a different screen, it is not a different system, it is in the console itself,” Jain said.
“This solution helped us digitize the knowledge of our experienced operators over many years so we could leverage it now and in the future,” Jain said. This is also just the beginning for Chevron and AI-based alarm guidance, which will serve as the foundation for the future of predictive and prescriptive operations.
