Digital transformation or just control system tuning?

AI will have an important role in improving process performance; however, the first steps should be to appraise the capabilities of existing technology.

There is no doubt that AI has the potential to transform the process and manufacturing industries. AI and related technologies such as IoT, edge computing and cloud-based data platforms have created performance and optimisation opportunities that would have been unimaginable just a few short years ago. Condition monitoring, operations management, process optimisation and fault analysis are just some of the potential uses for AI, a field which is still evolving and yet to reveal its full potential.

What has unfortunately been missed by new technology entrants spruiking AI is that many solutions to typical process problems have always been available through existing control system technologies. What is now wrapped up in the language of AI and presented as new and innovative solutions are often just solutions that have the means to be addressed through the technology already present onsite. In many cases being approached using AI tools that are not always appropriate for the problem – delivered by companies that do not fully understand process and manufacturing – has the potential to add a degree of confusion, which often does not solve the problem and in fact can lead you further from the source of the problem.

While not looking to discredit the potential of AI, if used in ways which are not complementary to existing technology it runs the risk of having a costly and negative outcome, which can make people disillusioned and mistrustful of the new technology, blinding them to the real benefits available when it is used appropriately.

What is AI?

So how do we decide what is appropriate use of existing or new technologies? First of all, it is important to understand what AI is. Artificial Intelligence is a set of technologies that enable computers to turn data into information and make decisions. At its core are a set of machine learning tools that find patterns in data and create models from these patterns. These tools include curve fitting of historical data (known as regression analysis) and statistical analysis.

This certainly has many useful applications, provided there is sufficient understanding of how these tools work and can be applied. Because they are trying to predict the future based on the past, without knowing anything about the processes they are modelling, they are likely to fail in the case of an unexpected event. Anyone who has monitored trends in the share market will understand how history is a poor indicator of the future. For example, if we create a model around the output size distribution of a ball mill based on input rock size, hardness and ball mill speed, then it can predict accurately up to the point other parameters not included in the model change, such as ball size. Once unexpected variability is introduced, the models developed can lose their potential impact.

What can be done with existing technologies?

Mathematical modelling is an extremely powerful tool that has been around for decades. Unlike AI, it does not simply look at the past to predict the future. Rather, the model is developed to simulate the actual process. Such models, when properly implemented, will accurately predict an output based on any kind of event. For example, a mathematical model of a ball mill will predict product size distribution for any kind of input event.

Advanced process control (APC) using multivariable control algorithms (MVC) has been used successfully across a range of industries over many years to provide tight control in highly interactive processes. APC is well established in other industries but its potential has been little understood in the mining industry. However, it has been shown to be highly effective in grinding circuit controls, providing considerable benefits in production and quality.

Other tools, such as mass and energy balancing and mathematical modelling have been around for many years and are not AI. Yet increasingly these tools are presented as new and innovative AI-based solutions. There are numerous examples of technology businesses that are new to the sector promoting virtual flow or level meters as AI. These are simple mass balance algorithms that can be easily implemented in existing site control systems, without the excess cost, complexity and potential for misdirection that can arise from not understanding the process or existing technologies.

There is a wide range of tools available using existing technologies to solve process problems. What is now called an ‘edge device’ is just a renaming of existing technology. Field control devices have been around for decades and are capable of performing the complex functionality claimed by edge devices. But the introduction of this term has opened up the control environment to the IT world, and this has its own problems. Whereas control systems companies have had many years of experience in developing robust devices that operate reliably in extreme industrial environments, using software platforms that are bullet-proof, IT-based companies are introducing edge devices that are not designed for the rigours of industrial processes and the reliability expectations of industrial customers.

There are applications for some of these new devices, but once again their use needs balancing against process requirements, applicability and whether they bring anything truly new and innovative to the mix. With existing field devices, it is possible to perform complex control and analytical functions — for example, real-time analysis of the noise in a pressure signal to detect pump cavitation. This is proven technology that adds the desired performance analysis and improvement, on a platform which is likely already resident on your site.

Where does AI fit into the picture?

New terms such as ‘digital twin’ are often used to refer to process models. But although the term is new, the technology is well established. So if these tools have always been available, why are there still problems to be solved, and what role does AI play?

Firstly, creating a mathematical model of a process or a piece of equipment, such as a ball mill, is more easily said than done. Fully understanding the physics of the ball mill is required to generate a useful model. But this is not easy to achieve. In cases such as these, building models using machine learning can provide a relatively simple way of building a useful representation that can be deployed quickly. These machine learning algorithms can be continuously evolving, reading data from the process and comparing it to the model. In this way, models can become quite accurate and informative over time.

Another example where AI can provide significant value is where there are networks of dependent processes, such as flotation cells and grinding circuits. Modelling these networks can be difficult to achieve, and AI can perform a valuable role in creating predictive representations of the process. One significant advantage of AI is that models can be created even when not all parameters are known or available. This is particularly valuable with processes such as flotation where critical parameters cannot be measured or inferred. In this case, building statistical models based on a large amount of accumulated data can assist in predicting production of the network of cells.

One of the challenges of introducing AI into a process environment is that of IT/OT convergence. This convergence is a result of the need to exchange data in real time between the process and analytics systems that are often cloud based. This introduces a range of issues around system availability and security, and requires a set of skills that are shared between IT and OT. Therefore, the need for these teams to work together is now a necessity, and requires organisational changes within businesses. IT/OT convergence also provides opportunities for tighter integration between control and IT systems, enabling seamless integration between process and business information.

Conclusion

AI has an important role to play in improving process performance and Yokogawa are investing heavily in this space with solutions already on the market. However, when starting the journey of technology evaluation to resolve issues with AI in mind, the first steps should be to appraise existing technology. Technology already available onsite can provide solutions to typical process problems, designed specifically for the process, without the cost and having to rely on complex black-box analytics. Solutions are often already residing on the plant, with process engineers who understand the technology, just waiting for someone to implement them.

Apart from the cost of unnecessary deployment, inappropriately implemented AI solutions can fall short of expectations, meaning that people are less likely to seek out AI technologies when they would be the right choice. There are many such instances where poor understanding and deployment of new technologies — such as fieldbus and wireless — have caused negative press around highly advantageous new methods for improving processes.

When considering solutions to a problem, the first question you need to ask is: are the physical principles of the process understood? Can information be inferred from other data? If so, generally speaking, the problem can be resolved with a simple algorithm in an existing system. If not, AI may be the best option, and this may open up new possibilities for your process. Before embarking on such a project, it is always worth talking to a range of vendors and engineers with expertise in the area to determine this best approach. This has always been the case, but even more so with the range of complex technologies available on the market.

In summary, when embarking on this journey, these steps will help to ensure that you choose the right technologies for the problem:

• Clearly define the nature of the problem. Is it a physical problem, such as difficulties in measuring an important parameter? Is it a control problem where process stability is difficult to achieve?
• Talk to the process engineers about the principals of the process. Are they understood? Can they be modelled?
• Consider the existing technologies available onsite. Talk to your OT people and the technology vendors about how the systems can be used to solve the problem.
• If exitsing systems cannot be used as a solution, consider an AI solution. This will involve site OT engineers, corporate IT and IT/OT specialists who understand the overall requirements, including the logistics of connecting the control system to the corporate network and the cloud.
• Consult with the control system vendor who has the OT/IT capability to facilitate the implementation and assist with site collaboration to achieve the best solution.

*David Walker is a chemical engineer with over 30 years’ experience in control systems across a wide range of technologies and industries, including oil and gas, mining, power, water and waste and chemicals. As Chief Engineer for Yokogawa Australia, he provides technology support, project execution management, engineering improvements and standardisation.

Top image: ©stock.adobe.com/au/masterart2680