By Jennifer Shine, Principal Solution Engineer, Imubit
Industrial AI is getting a lot of attention, and for good reason. Across the process industries, companies are investing in digitalization on the expectation that data and algorithms can raise plant performance. If you have spent time trying to make sense of the market, though, you have likely come away more confused than confident.
LNS Research published the blog "Five Ways Industrial AI is Shaking Up Manufacturing", which breaks the broad industrial AI market into more manageable, use-case-driven categories: asset monitoring, process analytics, process control, productivity and safety. Most of the vendors listed capture value through advisory insights that a person must act on. A few vendors in the process control category provide full closed-loop automation of the insights they identify. That gap says a lot about the market.
Most AI providers still work in open-loop advisory mode, offering dashboards and insight but stopping short of optimized control, because closed loop is hard.
Why closed loop is hard
What keeps companies from closed-loop AI is the complexity of the task. Industrial plants are secure, high-stakes environments, and any AI system that closes the loop, making real-time decisions that affect throughput, emissions or safety, has to meet a high bar.
The work goes well beyond data science. A closed-loop system has to integrate with distributed control systems (DCS), programmable logic controllers (PLCs), existing model predictive controllers (MPC) or legacy advanced process control (APC), or any combination of them. It has to do that within strict cybersecurity protocols, while meeting rigorous process safety requirements and showing deterministic behavior to some of the most technically demanding and skeptical stakeholders in the industry. Change management and operational readiness come after all of that.
Most AI vendors stop short of closed-loop control because they do not have the technology stack, or the domain knowledge, to cross that threshold. They may offer strong analytics or a well-designed user experience, but when it comes to turning those insights into action in the plant, they hit a wall.
Advisory-only AI leaves value on the table
The question to start with, when judging how much value an AI product can deliver, is whether it depends on a person to act. Domain experts are essential. They understand the plant, its context and its goals in a way no model can. But suppose we can build an AI model that accurately predicts and prescribes what should be done to best optimize the plant, and acting on it requires the console operator to move a dozen variables every five minutes. That is not a realistic ask, even of your best operator.
This is the trap of open-loop advisory systems. Their insights can be sophisticated, but they offer no reliable path to act on them consistently and continuously. The more complex the plant and its economic objectives, the more complex the optimal solution, and the more essential it becomes to automate execution. Capturing that value takes closed-loop optimization, where the model acts on what it recommends, inside limits set by plant engineers.
Closed loop in the plant
At Imubit, we work on the problems that keep other vendors in advisory mode, in refining, chemicals and polymers, cement, and mining and minerals. Our Deep Learning Process Control® (DLPC®) technology was built for the demands of the process industries. It integrates directly with plant control systems, operates securely on-premise within the plant's industrial network, and continuously optimizes entire units.
A Deep Learning Predictor learns a nonlinear dynamic model of the unit from its own operating history, guided by first principles, so it captures how the unit behaves, including where traditional models break down. A Deep Learning Controller, trained by reinforcement learning across millions of simulations, learns the control policy that moves the unit toward its economic objectives. Engineers configure the application and can audit it, and it is designed to keep adapting in a live plant.
The controller writes targets to the APC or DCS minute by minute, inside limits set by plant engineers. In a fraction of the time it would take an insight to travel down the communication chain to the control room, the adjustment has been made. The controller executes the optimized strategy and keeps adjusting as conditions change, so the value is captured instead of waiting on an emailed target or a number on a screen.
Adoption and trust matter as much as the technology. A team of former process engineers and process control engineers guides each client through a structured delivery process and involves plant engineers and operators at every step. Imubit has over 100 closed-loop controllers deployed worldwide, the largest install base in the industry.
What to ask a vendor
The industrial AI market will stay noisy, but it does not have to be confusing. When you evaluate vendors, look past the dashboards and ask whether the product delivers impact at the level of the plant, and whether it can act on its own recommendations in closed loop, inside the envelope your engineers set.
