Reinforcement Learning? Industrial AI? Closed Loop AI Optimization? Imubit webinars spark thought provoking questions from industry professionals like you.

By Allison Buenemann, Product Marketing Manager at Imubit

Audience participation was at an all-time high in our most recent webinar featuring Imubit customer Big West Oil. With a record number of audience questions raised, we decided to make the answers more accessible by assembling our most frequently asked questions into this blog post.

• Product Architecture FAQs
• Application / Case Study FAQs
• Continuous Improvement FAQs
• Technology Comparison FAQs
• People FAQs

Product & Architecture FAQs

Q: Can you elaborate on Imubit’s architecture and how the software is able to share data with everyone?
A: Imubit’s architecture consists of two main components 1) the cloud-based Industrial AI platform (which provides a collaborative environment for members of all teams to build, evaluate, and track performance of the closed loop models), and 2) the on-premises closed loop Deep Learning Process Control application. There are two on-premises deployment options for DLPC in the DMZ and PCN. There are no two-way exchanges of data between the on-premises controller and the cloud database. The on-premises closed loop optimization communicates with numerous on-prem sources, including the DCS, laboratory, and data historians, utilizing OPC/OPC-UA an industry standard protocol.

Q: Is the runtime control running on the process control network (PCN)?
A: Imubit’s on-prem DLPC application can connect to either the DCS or to existing APC. The software sits on the PCN level in order to communicate with these systems. Data extraction for model training from the site historian is typically done from the business network.

Q: How does Imubit integrate with other systems to get economic, pricing, or cost data for the optimization?
A: Some plants will push prices from their LP to their historian or straight to the DCS. When this is the case, Imubit can see in real time when there are changes to the model based on updates to the price deck. Some customers prefer to have more control over when these updates occur and in these cases, they will manually enter price changes into the on-prem DLPC application. In order for models to effectively control/optimize need up to date prices… So the more often you can input changes to prices, the better a job Imubit can do at optimizing!

Q: How often do you collect data from the plant and how often do you update set points?
A: This is a configurable parameter, but typically data is collected every 3 to 5 minutes. The range for all Imubit applications spans from a 1 minute read/write frequency to up to an hour.

Q: Does Imubit DLPC have automated feedback (lab results, analyzer results) to help adjust inferentials without human intervention so the model stays “up to date” with slow changes to unit operation or feedstock quality?
A: Imubit biases inferentials onsite using cumulative sum method or straight linear biasing method. This is something that the Implementation Engineering team works with the onsite team to do. This is very similar to how other technologies bias their inferrentials to prioritize most recent lab results.

Application FAQs

Q: Has Imubit worked with other refineries to solve similar optimization problems [around FCC units]?
A: Imubit has >70 live DLPC applications running within our global customer base (as of this 2Q2024 publish date), and more than 20 of these are on FCC units. These applications span optimization strategies on all stages of FCC operation including preheat, fractionation, reactor/regen, C3/C4 splitter, gas plant, and others.

Q: Can you provide more details on the application of Imubit DLPC for FCC units?
A: Imubit has extensive experience working on optimization strategies for FCCs as highlighted above. For a full example, download our FCC case study.

Q: Has Imubit implemented DLPC in ethylene plants?
A: Imubit has experience working on optimization strategies for C2 and C3 splitters and steamcracker furnaces, and we’re always looking to explore new ways to apply the technology! If you have an idea for how Imubit’s closed loop AI could help optimize your unit we would love to schedule an introductory demo call to hear about it.

Q: Can you elaborate on how you would handle solving multi-objective optimization problems?
A: Imubit models can be trained using multiple objectives. In training, tunable parameters are introduced, so that in real time, the user can update the relative importance of each objective.

Q: For FCCs with varied feed composition, does the DLPC need to take that into account?
A: Yes, the model definitely needs to, and does, take these different operating modes into account. For example, Imubit customer Big West Oil mentioned having 3-4 base crudes in varying percentages. The key is to ensure your training data set contains enough operating data for each of the different feed composition regimes for the reinforcement learning-trained controller to know how to respond optimally to the hundreds of millions of different potential scenarios that could be encountered during real time operations.

Continuous Improvement FAQs

Q: Do you have any examples of how your application is able to work around instrument limitations (for example, minimally instrumented sites, or an MV or CV outage)?
A: If it’s unplanned, the model will turn itself off. There can be contingencies built into the model for planned MV outages, for example an instrument on a line that’s only run through a certain percentage of the time. These are called flexible or flex MVs. With CVs, can have flexible constraints if you are only controlling to a constraint some of the time. Some other features that you can use include bad value protection along with instructions of what to do in case of values in that range.

Q: How much input does a DPLC need from a process or process control engineer? Are the Deep Learning models manipulable? If yes, at which degree?
A: Process and Process Control Engineers are brought into the modeling process to incorporate their domain expertise. The deep learning models aren’t magic, and we can’t just throw data at them. We need the domain experts to define the “rules of the game” – i.e. which variables are independent / dependent, the economic objective, the time to steady-state, constraints, etc. We let the model learn as much as possible from the data, but when important information isn’t perfectly contained within the data, then we leverage the knowledge of your experts to incorporate information like mass balance or gain tuning. We use Automated Machine Learning (AutoML), to translate this domain expertise into data science parameters.

Q: How long would it take to retrain the DLPC model on a new operating regime? For example, in the webinar with Big West where their crude unit was updated during a turnaround?
A: The rate limiting step in retraining a DLPC model on a new operating regime is collecting enough data in that regime to represent a broad spectrum of operations. This typically means collecting a minimum of a couple of months worth of operating data. Once the data is collected the time for the retrains to be executed is typically two weeks or less.

Q: How often must the model(s) be retrained, and which plant personnel are involved (Process engineers, process control/DCS engineers, economics and planning, etc.)?
A: Beyond the initial model training, subsequent retrains really do not require too much input from the plant. Because Imubit is collecting plant data continuously, there is no burden on the customer/plant to export data. As part of the continuous improvement workflow, the Imubit team is continuously monitoring customer’s DLPC applications, and being proactive about suggesting retrains when they see shifts in operation. The Imubit team is able to do the majority of the model retrain process without requiring customer assistance, though we like to bring in customers to review updates to the model following a retrain.

Technology Comparison FAQs

Q: What are the advantages of using Imubit DLPC over any other traditional APC package?
A: While there are a lot of technical differences between the way Imubit DLPC works compared to traditional APC, a couple of highlights include 1) the way that the dynamic process simulator is trained using a neural network (rather than first principles), and 2) the use of Reinforcement Learning (RL) to train and pre-solve for all possible optimization solutions in an offline cloud environment (no online solver).

Training the process simulator with a data-first approach frees us of the limiting linearizations, approximations, and assumptions of first principles-based models. The use of RL to do offline solving of all possible scenarios is unique to Imubit. RL is an incredibly strong and powerful technology because of the scale of the training that is capable. It quite literally learns similar to how a human brain learns. For example, if you’re learning to walk as a child, you learn through trial and error. The APC analogy of walking is that every step you take you get out your protractor/ruler/calculator, measure the gravitational force and you calculate in real-time, how high you need to lift your foot to take a step.

From Big West Oil – After evaluating APC many times it felt like it would be resource intensive and we might be limited in the results delivered by our limited instrumentation. Imubit felt like the next innovation in the closed loop optimization and control world. It was something that would work even with our limited instrumentation, and where a lot of the supporting work to implement and maintain would be done by the Imubit team.

People FAQs

Q: [to Imubit customer] How difficult has it been to sell operations teams at Big West Oil on AI-based process control?
A: At first, operations were skeptical. They were skeptical of 1) whether the technology would work, and 2) if it would put them out of a job. Starting with a simple model for DLPC helped a lot in building operator trust. By focusing on the debutanizer tower, which had only two MV handles which they were familiar with and knew how to interpret, they were able to see the results and understand them. Building this initial trust is critical to moving on to more complex control paradigms where explainability gets trickier.