Artificial Intelligence for Autonomous Control Elucidated at ARC Industry Forum Asia

By Sharada Prahladrao

Project Success Story

At the online annual ARC Industry Forum Asia titled Accelerating Industrial Digital Transformation and Sustainability on July 12-14, 2022, Yokogawa participated as a Gold Sponsor. This Forum saw registrations of 1,600+ delegates for the two language tracks – Japanese and English. In the session on AI and Machine Learning, Dr. Darius Ngo, Senior Vice President, Head of Digital Enterprise Solution, Yokogawa Engineering Asia, shared an implementation of Artificial Intelligence (AI) for autonomous control based on an actual plan using reinforcement learning – Factorial Kernel Dynamic Policy Programming (FKDPP). FKDPP is a disruptive innovation that allows for a different dimension of control. This AI technology can be applied in the energy, materials, pharmaceuticals, and many other industries.

At the end of this session, Dr. Ngo joined the other speakers for the panel discussion. This blog captures the key points of  Dr. Ngo’s presentation and his views at the panel discussion. The entire session can be watched on YouTube.

Challenges of Process Industries

Process industries (oil refineries, petrochemicals, steel, water etc.) require complex control of temperature, pressure, and flow rate due to chemical reactions and other factors. Dr. Ngo explained this complex control scenario by giving the example of an oil refinery from refining to processing and final assembly. The 4Ms that impact quality and production are:

Artificial Intelligence

Manufacturers are now turning to explore cutting-edge technologies, such as AI and ML to autonomize operations. Since the launch of Industry 4.0 the focus area of AI has expanded. Dr. Ngo gave a diagrammatic representation of AI/ML in process control via a typically linear application over control layers of hierarchy. At Level 1 (sensor level) itself there is already an embedded AI; Level 2 deals more with the control layers – IoT network, DCS etc. At this stage AI can be embedded into a reinforcement learning card algorithm (FKDPP) on controllers. Level 3 and above leverage applications like visualization analytics  – AI embedded on image analytics on the field monitoring devices, robots and so on.  Above that there are a variety of applications and services for specific solutions; this is Yokogawa’s AI platform studio – Xperience and Responsible AI platform to create AI algorithms for specific applications.     

Factorial Kernel Dynamic Policy Programming (FKDPP)

The FKDPP algorithm was jointly developed by Yokogawa and the Nara Institute of Science and Technology (NAIST) in 2018. It was recognized at an IEEE International Conference on Automation Science and Engineering as being the first reinforcement learning-based AI in the world that can be utilized in plant management. FKDPP was run on a simulator of a vinyl acetate manufacturing plant and it operated the valves to maximize the volume of products while ensuring adherence to the quality and safety standards. Stable and optimized valve operation was achieved in 30 learning trials.

FKDPP employs a factorial policy model and a factor-wise kernel-based smooth policy update by regularization with the Kullback-Leibler divergence between the current and updated policies. Compared with previous methods that cannot directly process a huge number of actions, Yokogawa’s proposed method leverages the same number of training samples and achieves a better control strategy for vinyl acetate monomer (VAM) yield, quality, and plant stability.

Salient Features of FKDPP

  • Can be applied to most kinds of control
  • Increases productivity
  • Simple
  • Explainable operation
  • Safety levels same as conventional systems

Case Study

In 2019, Yokogawa Engineering Asia succeeded in an experiment using a control training device. A three-tank level control system was set up via a laptop PC. Although the system can be controlled with conventional PID technology, it was shown that FKDPP can reduce the settling time by 50-70 percent, while also preventing overshooting and maintaining the tank’s water level. This was demonstrated by a video that showed the differences between the 1st, 20th, 25th, and 30th iterations of reinforcement learning based AI (FKDPP algorithm). The three basic steps from FKDPP model generation to actual control are: target setting, constructing the AI control model, and AI autonomous control.

Over the last three years the efficacy of the FKDPP algorithm has been tested and projects were started with ENEOS Materials Corporation and NTT DOCOMO. Next, Dr. Ngo spoke about how FKDPP balanced quality and energy savings. The media opined that FKDPP “can greatly contribute to the autonomization of production, ROI maximization, and environmental sustainability.”

Future Outlook

In this context, Dr. Ngo spoke about Yokogawa’s vision of Industrial Automation to Industrial Autonomy (IA2IA). A survey of 534 decision makers at 390 manufacturing plants reveals that 42 percent believe that in the next three years the application of AI to plant process optimization will have a significant impact on industrial autonomy. The envisioned application of 5G, cloud, and AI for industrial autonomy will enable optimum control at anytime and from anywhere.


Summarized below are Dr. Ngo’s responses during the panel discussion.

Is the design suitable for different interfaces?

Presently the implementation is through OPC interface; but in the long run the company will integrate full visualization. The multi-vendor data will be pulled from a data lake and put into the system.

Why was the pilot project on the chemical plant limited to 35 days?

For this chemical plant there was a routine maintenance on the 36th day that’s why it was stopped on the 35th day. After that when the plant was restarted it was on AI control.

What is the time taken for FKDPP to learn operator actions before putting in autonomous control?

Safety is always the key pivot of Yokogawa’s implementations. FKDPP learning was from plant historian, including operator action simulation, to ensure a safe autonomous control. The duration of the learning depends on the complexity of the control. In this particular plant the time taken was short because of the adaptive processes implemented in FKDPP.

Going forward do you see AI replacing traditional PID?

We are more interested in addressing what PID and APC cannot do and fill those gaps and improvise by using AI on that. However, in the future this may happen. Even the academia is trying to push the ideal of all-encompassing AI, but the intelligence should be based on the fundamentals of PID kind of methodology. This is a period of transition – even the academia will take time to adjust to AI as a control strategy rather than fundamental engineering.



Engage with ARC Advisory Group