An AI-based autonomous control runs the HVAC system of a semiconductor factory. This feature originally appeared in the October 2022 issue of InTech magazine.
As process control technologies advance, one concept that is gaining prominence is autonomy. Unlike conventional automation, one of the main differentiators of autonomy is the application of artificial intelligence (AI) so that an automation system can learn about the process and make its own operational improvements. . Although many companies find this idea intriguing, there is understandable skepticism. The system’s capability is only as good as its core algorithms, and many would-be users want to see AI working elsewhere before they wholeheartedly buy into the idea. These concrete examples are beginning to emerge.
Yokogawa’s autonomous control systems are built around Kernel Factorial Dynamic Policy Programming (FKDPP), which is an AI reinforcement learning algorithm first developed in a joint project of Yokogawa and the Nara Institute of Science and Technology (NAIST) in 2018. Reinforcement learning techniques have been used successfully in computer games, but extending this methodology to process control has been difficult. It can take millions, or even billions, of trial-and-error cycles for software to completely learn a new task.
Since its introduction, FKDPP has been refined and improved for industrial automation systems, typically working with factory simulation platforms used for operator training and other purposes. Yokogawa and two other companies have created a simulation of a vinyl acetate manufacturing plant. The process required modulating four valves based on input from nine sensors to maximize the volume of product produced, while meeting quality and safety standards. FKDPP achieved optimized operation with only about 30 cycles of trial and error – a significant achievement.
This project was presented at the IEEE international conference in August 2018. In 2020, this technology was able to control the entire process manufacturing facilities, but on highly sophisticated simulators. So the next question became, is the FKDPP ready for the real world?
From simulation to reality
Yokogawa answered this question at its Komagane semiconductor factory in Miyada-mura, Japan. (Figure 1). Here, much of the production takes place in clean room environments under tightly controlled temperature and humidity conditions necessary to produce defect-free products. The task of the AI system is to operate the heating, ventilation and air conditioning (HVAC) systems optimally by maintaining the required environmental conditions while minimizing energy consumption.
It is understandable that a real application selected for this type of experiment would be modest in scale with minimal potential for security risks. This conservative approach may be less dramatic than that of an oil refinery, but that does not reduce its validity as a proof of concept.
At first glance, running an HVAC system on its own might not seem complex. But HVAC systems supporting the tightly controlled clean room environment represent 30% of the total energy consumed by the facility, and therefore represent a significant cost. Japan’s climate varies through the seasons, so there are adjustments needed at different times of the year to balance heating and cooling, while maintaining humidity control.
The facility resides in a mountain valley at an elevation of 646 meters (2,119 ft). It has a temperate climate and tends to be relatively cool, with an annual temperature between -9° and 25°C (15.8° and 77°F). The factory produces pressure sensors based on semiconductors (Figure 2) that fall into the company’s DPharp family of pressure transmitters, so maintaining uninterrupted production is essential. Even though this demonstration takes place in one of Yokogawa’s own factories, the cost and production risks are no less real than those of an external customer.
The installation location is outside the local natural gas distribution system, so liquefied petroleum gas (LPG) must be brought in to provide steam for heating and humidification. Air-cooling operates on conventional mains-supplied electrical power. The two systems work together when necessary to maintain critical humidity levels.
Complex power distribution
Considerations surrounding energy consumption in Japanese manufacturing plants begin with the high domestic cost. Energy in all its forms is expensive by global standards, and efficiency is paramount. The Komagane facility uses electric furnaces for processing silicon wafers, and there is a need to recover as much waste heat as possible from these operations, especially during the winter months.
To be considered a success, the autonomous control system must balance many critical objectives, some of which are mutually exclusive. These goals include:
- Strict standards of temperature and humidity in the clean room environment must be maintained in the interest of product quality, but with the lowest possible consumption of LPG and electricity.
- Weather conditions can change dramatically over a short period of time, requiring compensation.
- The clean room environment is very large, so there is a high degree of thermal inertia. Therefore, it may take a long time to change the temperature.
- The cleanroom equipment also provides heat, but this cannot be regulated by the automated control system.
- Waste heat from electric furnaces is used as a heat source instead of propane gas, but the amount available is highly variable, depending on the number of production lines in use at any given time.
- The heated coolant from the boiler is the main source of heat for the outside air. If more heat is required than is available from this recovered source, it must come from the LPG burning boiler.
- Outside air is heated or cooled depending on the local temperature, typically between 3° and 28°C (37.4° and 82.4°F). For most of the year, the outside air needs to be heated.
The existing control strategy (Figure 3) is more complex than it appears at first glance. Beneath the surface, the mechanisms involved are interconnected in a way that has changed over the years as plant engineers have worked to increase efficiency.
There have been many previous attempts to reduce LPG consumption without making large investments in new equipment. These incremental improvements reached their practical limits in 2019, leading to the implementation of the new FKDPP-based control strategy in early 2020.
The implementation team chose a slow day during a scheduled production shutdown to commission the new control system. During this day, the AI system was allowed to do its own experimentation with the equipment to learn its characteristics. After about 20 iterations, the AI system had developed a process model capable of running the entire HVAC system well enough to support real-world production.
Over the weeks and months of 2020, the AI system continued to refine its model, making routine adjustments to accommodate changes in production volumes and seasonal variations in temperature. The ultimate benefit of the new FKDPP-based system was a reduction in LPG consumption of 3.6% after its implementation in 2020, fully based on the new AI strategy, with no major investment required.
FKDPP-based AI is one of the key technologies supporting Yokogawa’s transition from industrial automation to industrial autonomy (IA2IA), complementing conventional proportional-integral-derivative and advanced process control concepts in many situations, and even replacing complex manual operations in other cases. Real-time control using AI reinforcement learning, as demonstrated here, is the next generation of control technology, and it can be used with virtually any manufacturing process to bring it closer to fully autonomous operation.
All images courtesy of Yokogawa
This feature originally appeared in the October 2022 issue of InTech magazine. You can read it here: https://www.isa.org/intech-home/2022/october-2022
Read InTech magazine
Did you enjoy this great article?
Check out our free e-newsletters to read other interesting articles.
#Case #Study #AIBased #Autonomous #Control