Process Safety: AI and Machine Learning Transform Compliance, But Humans Still Key

In this episode,  Trish and Traci welcome guest Dheerajkumar Narang, whose research examines how AI and machine learning can enhance process safety compliance. Traditional compliance methods are time-consuming and fragmented across different systems, while AI can automate data collection, identify leading indicators and predict compliance outcomes. Key challenges include system integration with legacy infrastructure and maintaining domain expertise for regulatory updates. AI should streamline processes to allow operators to focus on critical tasks.

Transcript

Welcome to Process Safety with Trish and Traci, the podcast that aims to share insights from past incidents to help avoid future events. Please subscribe to this free award-winning podcast on your favorite platform, so you can continue learning with Trish and me in this series. I'm Traci Purdum, Editor-in-Chief of Chemical Processing, and joining me as always is Trish Kerin, director of Lead Like Kerin. Hey, Trish, this is our first episode since winning the Jesse H. Neal Award for best podcast. The Neal Awards are a pretty big deal in the business media, so it's very exciting. So congratulations to you.

Trish:  And congratulations to you, Traci. I noticed you just added award-winning podcast, and it brought a smile to my face. Yeah, pretty excited that we've been recognized by such a prestigious award, and I think it just really highlights to us how much people perhaps appreciate what we say. And as the article said in Chemical Processing, we do get feedback from all over the world of people that listen to our podcast and really love it, and it's just great to be recognized for that, I think.

Traci: Indeed, it is. It was very exciting news, very validating news for us, and we're on this path for process safety. So, thanks for being on the path with me, and today we're bringing somebody else with us. We're welcoming a guest to our podcast. Dheerajkumar Narang is a corrosion and reliability engineering specialist, process safety activist and digital innovation researcher. His professional and research interests are focused on critical infrastructure, reliability, smart asset management, e-governance and industry 4.0 solutions in the process plant industries, oil and gas, petrochemical, fertilizer, power generation, and green energy. His professional career includes experience in developing, managing and implementing reliability engineering programs for numerous global energy companies across the United States, India and the Middle East. He has also published and presented several novel methodologies, as well as served on many process safety conference committees. Welcome, Dheeraj.

Dheeraj: Thank you so much for having me, and congratulations to both of you for the Neal Award.

Trish: Thank you.

Can AI Bolster Process Safety and Compliance?

Traci: Well, thank you very much for joining us. We certainly identify with the process safety activist title. In today's episode, we'll discuss how artificial intelligence and digital technologies can bolster process safety practices and compliance. You've done some extensive research in this area. Can you briefly tell us about that research?

Dheeraj: Yes, sure. So, my research actually began with one of my project assignments back in India with L&T Energy, and it was about an e-governance solution. I tapped into the world of digital solutions and I took online courses, and that's how I got interested in building machine learning for asset integrity and process safety management. So, when I came to the US for my PhD, I started researching in this area, and I found that there is a tremendous opportunity to digitize the entire regulatory compliance process, right from the process plant end to regulatory authorities. So I have done a few modeling projects as part of my PhD. Some papers are published, and some papers are in the pipeline. So, mostly it's about how we can automate, reduce the economic burden on process plants, and regulatory authorities, and then we can have a PSM-compliant industry.

Traci: Well, it was very interesting, and you had sent along some information beforehand that I was able to research and look into, and I wanted to have you, based on your research, tell us a little bit about some of the challenges with traditional non-digital methods of safety compliance.

Challenges with Non-Digital Methods

Dheeraj: Yeah, so with a non-digital method, let's take an example. Say there is a process plant facility, and they want to achieve PSM compliance. Now they will need to collect all the data for all the elements. Those are relevant to OSHA, PSM or, depending on where the process plan facility is located, they would have their own regulatory compliance procedure. Now it takes a lot of manual effort and time, and man-hours have an enterprise-wide view of the entire process safety management data. So, for example, you will have incident investigation reports in a differential software, and you will have your MOC process in a differential software. So, how do we integrate all this and have a unified digital solution, which could help process plan facilities to achieve compliance at a minimal cost, and at the same time find out the indicators and what are the causes effects of the incidents, if any at all in the past?

Traci:  Trish, what are your thoughts on some of the challenges with the traditional methods of safety compliance?

Trish:  Yeah, Dheeraj, I think you are absolutely correct in that, and no matter where you are in the world, the compliance burden of assembling all of the information and even at that high level, doing some basic analysis on it is a massive task. And so, I think if there is better ways that we can use machine learning to basically collect all of this information for us and be able to sort it in a way that we can make good engineering judgements on it, and those aided with the use of AI as well, I think it's a fantastic direction and I think it is the way of the future. I think it is a thing that we need to get on board with, and there are a lot of people who are very nervous about this idea of machine learning or AI, and we need to still be doing the risk assessment.

Yeah, there are elements of the risk assessment that we do need to be doing, absolutely. But the information that goes into that, I think, is where we can make an enormous difference in some of these technologies and really push toward that. And I agree with what you said as part of your research as well. This can have enormous positive implications for the regulators and the authorities as well. And to get good process safety outcomes, not only do we need facilities operating well, but we also need a capable and resourced regulator. And if we can regulate in more effective ways, that has an enormous benefit too.

Machine Learning and Leading Indicators

Traci: Trish, you mentioned making a difference in process safety and streamlining things. Dheeraj, I want to understand a little bit more about how machine learning algorithms might identify leading indicators of process safety incidents that traditional compliance methods might miss.

Dheeraj: Yeah, so before giving an answer to that question, I just want to highlight subdomains of machine learning and how it might relate to different aspects of regulatory compliance. So, machine learning is a subset of AI. What AI does, artificial intelligence, is they mimic human behavior or they aid machines in mimicking human behavior. Now it could be as simple as predicting a particular class, saying a yes or no, or it could be as complex as a large language model. So, there is a spectrum of machine learning and AI domains that would be applicable to different aspects of regulatory compliance.

So, I wanted to just mention that for the benefit of the audience and the indicators. So, finding the indicators for a particular facility in the domain of PSM compliance, one can use unsupervised machine learning. So, what unsupervised machine learning is capable of doing is, you have all the scattered data, right? You have all your data, say let's pick an element, process safety information; you have all your data, how do you compile data and find out the leading indicators that would be relevant to judge whether you have achieved compliance against that particular element? So, in that case, unsupervised machine learning methods such as principal component analysis would be useful, and it could give you different correlations and features that are correlated to one another, and that's how one can identify the indicators.

Trish: Yeah, I think that's a really great example, actually. And one of the ones that has often been challenging, particularly around looking at leading indicators, because I mean, one of the issues that I think we have is even when you have very solid data correlations, it is still not absolute proof that something is going to happen. It is more likely to happen. And I think that's where we need to really make sure that people understand that leading indicators are critically important. They're going to point you in a direction of where you need to focus your attention, so that you can resolve an issue or prevent something from occurring, but they're not absolutely predictive in that we can't guarantee it's going to lead to, A is going to lead to D.

What we do know is that A will more likely lead to D, with an understanding of those correlations. So that's a really important distinction. Sometimes I think people get confused with lead indicators and think they're predictive. What they do is they give you an opportunity to manage your future, because you can see you're more likely to have this occur. So for example, you might have a corrosion rate, which you're a corrosion specialist, a corrosion rate on something that you're measuring and you're monitoring, and you can infer from that data, that particular piece of equipment is likely to reach the end of its serviceable life at a certain point.

You can't predict it to the hour, typically. What you can do is have a range of we are going to lose enough material that we're going to lose our serviceability of this piece of equipment at a certain point in time and we can get a good range of that, but other things could happen and all of a sudden that corrosion might accelerate in some way, and it happens earlier. So I think we need to really understand the correlations we're finding and use them to inform our decisions. And I think that's the really key part about what we should be doing with any leading indicators. Leading indicators are there to inform our future decisions and make sure we make good decisions.

Dheeraj: So, I completely agree with Trish. So, say we have the leading indicators. Now these are approximations since it is the unsupervised machine learning, you will get clusters of reasoning like this particular feature correlates with this much, like an amount, like 70%, 80%. But then, what one can do is one can take those indicators which are more relevant, which are more correlated in terms of quantitative technique language, and then model those indicators as independent variables, and then you can have a dependent variable predicting whether you will achieve compliance or not. So it is possible to build different domains of machine learning on top of each other to make the collective model work for you. When I say you, I mean the facility.

Trish: Yeah, absolutely agree with you. Absolutely.

Machine Learning Implementation

Traci: Dheeraj, can you walk us through a machine learning implementation life cycle at a process plant facility? I know that you had some examples of that in your research.

Dheeraj: Yeah, sure. So, it's going to be difficult to touch all aspects of process safety management compliance. I'm going to pick one element. Let's say, and I'm going to pick my favorite mechanical integrity, because I'm biased. I have a mechanical engineering degree.

Trish: I love that.

Dheeraj: So, let's say we have a mechanical integrity element and we want to find out, and implement a machine learning framework to predict whether we are going to hit, we are going to be compliant or not. The first step would be to collect the data. Now, when I collect the data, I'm going to find in different types of formats. I'm going to be finding a U1A form, as they file for pressure vessels in the US, which is like 1940s. So I have just a scan copy of it, super old. And I will also find super nice digital records, which are for equipment which are new. Now, how do I compile all this and make it work for my machine learning framework?

So, the very first step would be data pre-processing, that's where I will define features, and I will convert all the categorical data into some sort of ordinary data. So it'll have some meaning to it, like low, medium, high, or hydrocarbon, non-hydrocarbon, if it's a fluid containment. And then I will convert those ordinary variables into discrete variables or continuous variables, depending on the properties that you know how one can pre-process them. And when I get the numbers, then it is just a matter of effective modeling, and I can have a predictive model, a supervised machine learning model saying yes or no, whether you're going to hit a compliance or not. Now, to find out which feature to model to begin with, and to collect the data. That's where the unsupervised machine learning, the principal component analysis, as the name suggests, I'm going to find out the principal components relevant to my study.

But to find out the principal components, you also need subject matter expertise, right? So there's going to be a domain expertise involved as well. And in the end you have the model, you test it, you have the training data set, then you test it, you split the data set into two, you have the testing data set as well, and then you keep on testing and refine the model. The one thing I would say is year after year you have to keep on defining the model to have your accuracy, precision, and recall targets met. So that's like a nutshell of machine learning framework implementation for mechanical integrity.

Traci: Trish, do you have any questions there?

Trish: No, no questions. That sounds like a great structure for a model to really make sure that you're not only getting the data you need, but you're also refining the process as you go through. That iterative refining is critically important to continue to improve the integrity of the data that you're getting, and therefore your mechanical integrity as a result, too. I love it.

Keeping AI Up-To-Date

Traci:  And I think that's with the AI and machine learning is to continually tweak it to make it better and better and better. I think that's key for sure. Let's talk a little bit about process plant facilities and how they can ensure that the AI and machine learning systems remain properly validated and up to date with changing regulatory requirements without creating new compliance gaps. We've got a lot happening here in the US. How does it stay up to date?

Dheeraj: Yeah, that's a very good question. I think recently EPA, RMP regulations, they have I think two or three major requirements added there. But the PSM regulations, there have been some proposed changes ever since 2013 incident, but they haven't really done any final federal rulemaking for that. However, things could also change rapidly even within the aspects of implementation. Regulatory authorities can also interpret more stringently, and then facilities will need to tap up and meet the compliance requirements. So, this is where domain expertise comes into the picture.

We have the model. So when we do the feature selection and feature extraction, there is an important step of knowledge-based selection. So one feature selection aspect would be is going to be completely how the features will aid in statistical modeling/machine learning modeling. But one aspect is knowledge-based selections, where you don't lose the importance and relevance of the domain, because it's a process safety management domain. So it has to make sense in that way as well. So that's the only solution. Right now, there are no automated solutions for making sure your automated machine learning driven PMS and compliance are up-to-date. So that step is manual, and it should be manual.

Traci:  Trish, thoughts about that, the management of change, essentially for this?

Trish:  Yeah, I was about to add, as Dheeraj just said, it's a manual process and it should be manual, because we do need to make sure that we are keeping an oversight over what we are doing and making sure that we've got the right interpretation of any rule changes that occur, to build it back into our machine learning and help continue to train that. So I think to a certain extent it does need to continue to be manual as Dheeraj said, so that we can make sure we've got that right interpretation, make sure we've got the right information and make sure that we don't end up in a situation, where we just think that the AI is doing it all for us and we don't have to do anything anymore, because that's not what it's all about. It's about using these tools to streamline and to improve our business processes to such an extent that we can focus on other things that we need to be focusing on.

Traci:  I agree with that for sure. Let's now get into some of the organizational change management implementation strategies that you would recommend, Dheeraj, for facilities transitioning from traditional to AI-enhanced safety compliance systems. How does that work?

Dheeraj: That's a good question. So, just like any other IT/digital solution project, the number one bottleneck would be funding and the number two bottleneck, which could be facing the organizations, is the integration of the new systems into all the old ones. Sometimes people tend to miss that aspect, and that's the biggest challenge. So they can get the funding part, they see all the functionalities in the proof-of-concept presentations, they do a pilot, everything goes right, and they just give a go-ahead.

But they sometimes miss, I've seen it in my career, even without machine learning, like regular digital solutions, risk-based inspections or other mechanical integrity solutions, I've seen that they tend to miss the enterprise integration part. Like, how do you integrate your new digital solution into everything else that's already there, and are they even compatible? So there is a step of risk assessment that should be done by management and all the technical people, clearly stating where we can integrate it, where we cannot, and how much money it'll consume. And the stewardship of this solution, how much money that will consume and whether they're getting any benefit out of it. Those are the big challenges facing the facilities.

Trish:  In my experience, I absolutely agree, Dheeraj. The integration, I think really makes or breaks any of these new solutions. And I think the road past is littered with implementations that derailed in some way. My mixed metaphors there, it's just I'm on a road and I'm on a railway line. But absolutely, we so often underestimate how we're going to make that integration actually work, and we end up with situations of just proliferation of different systems all over the place. And then you get different parts of your business getting upset about something, and going and finding their own solution, because they can't make whatever you're doing work.

And we just end up with a digital mess in organizations. And if we could at some point, I don't know if someone out there can actually create something that talks to every system that already exists, that would just be my utopia. If I can go and get something off the shelf that's going to communicate with every one of my digital systems all in one interface, that would just be heaven. That would allow us to solve so many problems. But I guess that's a naive and hopeful answer.

Although I tell you, if there's anyone out there that can figure out how to do it, I think you'd be on a winner if you did, because so many companies are stuck with legacy systems and how do we interconnect them all together I think is a huge issue and we don't spend enough of our project focusing on that. We get distracted by other parts, as you said, Dheeraj, and we end up with this digital mess unfortunately. And I think that also then scares people off the idea of going down this path and we need to figure out how to get people on board, because this is the way of the future. This is what is going to allow us to focus on the more important things. We need machine learning to do the tasks that we don't have to do, so that we have time to do the things that we absolutely do have to do.

Dheeraj: Yep, I agree.

Traci:  And the machine learning and AI does those tasks that we don't want to do so much more thoroughly and really exposes things. So I think it's important.

Trish:  And faster. It does it so much faster than we can as humans compute.

Traci:  Dheeraj, I always toss out a question for Trish, and it's anything that you want to add that we have not touched upon that you think is important that we need to understand.

Dheeraj: I would say one thing. So we just discussed right about the integration part. Now, according to Moore's law, the computing power doubles every two to three years. So the digital solutions are going to only increase in terms of complexity and also usability. So to keep an industry, like a process plant industry, current with the digital solutions advancement is a very big task, and I think the entire industry needs to commit to that.

Traci:  Where does your research take you next?

Dheeraj: I'm most likely going to complete my PhD. That's what I'm doing. And then we'll see. I'll start looking for a job that is relevant to doing digital solutions for PS and compliance. That's the plan.

Traci:  Well, good, good. Trish, is there anything to add?

Trish:  I'm glad you brought in Moore's law there about computer power increasing as we go through, because I think that is a really important thing to remember, that we also need to be aware of the energy needed to continue digital solutions. And I'm suggesting we need to continue digital solutions, but we just also need to keep in mind that we need new and different energy sources to be able to maintain computational power, because the cloud takes a lot of energy. All of the networks that we have, the cooling systems that are required for servers, are absolutely massive, and we need to make sure that we understand we have to prepare for that, too. This is not just about using digital technology; it's about understanding where we're getting our energy sources and our future energy needs from to be able to allow this to occur.

Dheeraj: Yep, I agree. And one other aspect of that is recycling of the semiconductors, the chips, which are no longer in use in the data centers which support cloud. So there are end-to-end supplies and aspects we need to account as well. Yep.

Trish:  Absolutely. We actually need the resources in those chips. The resources are not endless. We are going to run out of rare earth metals at some point, so we need to be using all the ones that we've previously found and discarded. We need to figure out ways that we can actually recapture them and reuse that material, so that we can continue to go down this path.

Traci:  Well, Dheeraj, thank you for insight into your research and helping us to stay focused on the future. Trish, thank you for your insight on all of this. And unfortunate events happen all over the world, and we will be here to discuss and learn from them. Subscribe to this free award-winning podcast so you can stay on top of best practices. You can also visit us at chemicalprocessing.com for more tools and resources aimed at helping you run efficient and safe facilities. On behalf of Trish and Dheeraj, I'm Traci, and this is Process Safety with Trish and Traci. Thanks again, you too.

Dheeraj: Thank you so much.

Trish: Stay safe.