From ice-cream manufacture to steel making, automated control is a fundamental part of industrial processes. Control technology has been around since the industrial revolution, when centrifugal governors were used to regulate the speed of steam engines. But it was the advent of electrical, electronic, and computerized control technologies over the course of the 20th century that defined the field as we know it today. During that time, three major generations of control technology have evolved. Today, the emergence of an entirely new approach could transform the way companies control their assets.
Control 1.0: Making a difference with differentials
The first modern, flexible control technology was the programmable logic controller (PLC). Working on the proportional-integral-differential (PID) principle, a basic PLC monitors a single value, say the temperature of a freezer or a furnace, and compare that to a target setting for the machine. Depending on magnitude of the error, its duration and its rate of change, the PID controller output steers the machine, for example by adjusting a valve or changing the power flowing to a heating element or motor.
The earliest PLCs were extraordinarily simple devices by modern standards, but they could do some remarkable things. The PID control principle meant control engineers didn’t need to fully understand the complexities of the process under their care; with just three parameters to adjust, they could often achieve good performance with a little careful tuning. Using more elaborate configurations, such as cascaded control systems in which one PID controller generates the target value for another to meet, these systems could run quite complex processes. They are still widely used today.
Control 2.0: Adding the power of the PC
The second major generation of control system emerged with the development of personal-computer technology in the 1980s and 1990s. The greater speed and power of these machines allowed engineers to build control systems where numerous inputs and outputs were managed by a single computer, and to connect multiple machines together into networks. These distributed control system (DCS) designs provided a host of benefits for operators, not least the availability of clearer, easier user interfaces and the power to change setpoints from a distance. Under the hood, however, DCS systems ran on the same basic principles as their PLC predecessors.
Control 3.0: Accommodating complexity with advanced process controls
Toward the end of the 20th century, the availability of more powerful computers led a small number of the most demanding users to take a third step in control technology. These users were typically in industries running operations of huge scale and complexity, such as oil refineries or steel plants. For such applications, even small relative improvements in control-system performance could be worth millions of dollars in additional output, and owners were prepared to make big investments to achieve them.
The approach they adopted replaced the simple principles of PID control with complex and sophisticated mathematical models. By combining theoretical physicochemical models with carefully selected and calibrated sensor data, these advanced process control (APC) systems attempt to determine the current state of the process and decide how it should be adjusted to deliver the desired operating conditions. APC systems work very well, but their high cost means that their use is still limited to a minority of applications, in industries where plants are big enough and similar enough to pay back the enormous development effort required to implement them.
Control 4.0: Making ‘advanced’ process control ‘ordinary’ with intelligent automation
Today, a fourth generation of control systems is emerging, one the promises to exceed the performance of the APC approach at a fraction of its cost and complexity. This new approach uses advanced analytics (AA) or artificial-intelligence technologies (AI), such as so-called machine learning or even deep-learning approaches using artificial neural networks and equivalent methods.
These AA/AI systems work in a fundamentally different way from previous APC technologies. When IBM’s Deep Blue computer chess program beat grand master Gary Kasparov in 1997, it relied on thousands of explicit rules programmed by its designers, much like today’s industrial APC systems. 18 years later, when Google Deep Mind’s Alpha Go program defeated professional go player Lee Sedol, it used no such rules. Instead, the program developed its own strategy by analyzing past matches and playing thousands of simulated games. Alpha Go Zero, a more recent iteration of the company’s program, trained itself to beat its predecessor in three days, purely by playing games against itself.
It is now becoming possible to apply the same approach in industrial control systems, using an AI system that is “trained” using historical process data1. Many facilities have years of detailed records on operating conditions, process settings and the resulting performance. And once they are installed and operating, these systems can on learning, gradually improving their own performance over time.
For industrial companies, this new approach has some significant implications.
- AA/AI-based control works exceptionally well. Costly APC systems typically provide overall improvements in the order of two or three percent. AI systems can increase the performance of existing APC systems by an additional one or two percent and have boosted non-APC controlled systems by over 30 percent.
- AA/AI-based control doesn’t require system designers to model every detail of their process and build complex theoretical models—it learns those intricacies for itself. That means AI can be applied to complex processes with interactions that may not be well-understood. In real-world applications, AI tools have identified issues and improvement opportunities that eluded even experienced control engineers.
- AA/AI technology is much cheaper and easier to implement than earlier advanced control solutions. That paves the way for its use by companies with smaller plants or less common processes—especially as the latest generations of technology are user-friendly enough to be placed directly in the hands of the process engineer.
- AA/AI control approaches work best alongside people. Capturing the full potential of the approach almost always depends upon a combination of better process steering, technical upgrades to address issues and opportunities identified by the system, and improved performance management. Each of these improvement levers brings roughly one third of the total value at stake, and each requires the input of people with deep process knowledge.
Industrial companies are just beginning to exploit the potential of advanced analytics and artificial intelligence technologies in process control. That won’t be the end-game. Ultimately, the ability of AI to assist the control and optimization of complex systems will unlock entirely new ways to manage not just machines and production processes, but also entire businesses.
- It is a long-established feature of artificial intelligence development that, once a new technique is mastered, people stop calling it AI. As a result, many process engineers are already using AI technologies, albeit unwittingly. Advanced optimization techniques using decision trees or gradient boosting approaches all have their roots in AI research.