Adding AI and Machine Learning to Process Control
A new platform upgrade allows process control engineers and data scientists to collaborate on improvements for the plant.

By Steve Fiscor, Editor-in-Chief

A new software upgrade could improve the plant’s ability to maintain its setpoints.
Advanced process control (APC) systems at the mills and concentrators are enabling mining companies to optimize operations and improve recovery rates. These systems offer clear advantages for controlling the setpoints in the plant in real time, which is a game changer for feeds with varying grades. Their use in the plant has also helped identify other areas of possible improvement.

While dreams of a fully autonomous plant for the average operation might be just that, advancements in this area have been significant. APC systems today still require tuning and good maintenance programs that consider regular inspection of the sensors, devices and the loops that control them, along with training for technicians.

As computing speeds increase, system developers are hoping to create smarter systems that anticipate changes and make further refinements to the process using artificial intelligence (AI), data analytics and machine learning. One of the companies leading the way in this area is FLSmidth, which recently announced a major upgrade of its system.

Smarter Mineral Processing
Machine learning and AI are fast becoming a part of everyday life for the manufacturing sector and it’s expected to play a larger role for process optimization solutions in mining. To meet this growing trend, FLSmidth has launched the latest version of its APC solution, ECS/Process- Expert (PXP) V8.5 software, which now includes the ability to integrate new AI cognitive technologies and functions. “The new AI technologies are based on machine- and deep-learning algorithms that create their own understanding of a process by finding patterns in the raw process data. They then use this understanding to solve problems that help improve performance and sustainability,” said King Becerra, global product line manager for FLSmidth’s digital group.

As an example of AI at work, PXP V8.5 facilitates predictive modeling of process data signals that can be unreliable or unavailable. These are then used to create a more accurate model and controller of the actual process conditions. This improves the plant’s ability to maintain optimal setpoints and a faster response to conditions that may result in undesirable situations.

PXP V8.5 also incorporates new features to promote collaboration and cooperation. The new PXP Data-Books module, for example, allows the automation engineers and the data scientists at the mining operation to integrate their existing machine learning and deep learning algorithms into the PXP applications and control strategies. This function bridges the gap between the process control engineer and the data scientist, Becerra explained. “Often we find that the data scientist speaks one language, while the process control engineer speaks another,” Becerra said. “DataBooks allows these two professionals to work together in a more cohesive environment.”

While describing the PXP platform, Becerra emphasizes that it enables other applications as well. As an example, a mining operation may have a data scientist who is using deep learning to understand the particle size distribution of the cyclone overflow. Traditionally, that person would use a set of statistical functions. What FLSmidth has developed with the PXP is the possibility of using a machine learning algorithm that might solve this problem and easily integrate it into the APC.

Another new feature of PXP V8.5 is a range of sustainability dashboards that report the system’s performance in terms of environmental KPIs (key performance indicators). With these dashboards, operators can see how process improvements, such as more efficient use of energy, translate into lower CO2 emissions.

AI is already driving technological advances at manufacturing plants and Becerra believes that mineral processing operations could take advantage of these tools, too. However, he also cautioned that AI technologies have not achieved a level of development and maturity to replace everything currently used to control and optimize minerals processes. “We identify the problems we want to solve,” Becerra said. “We then apply the technology that is best suited to solve that problem, whether that be AI or something else. We think holistically — it’s not only about the technology, but also what is required in terms of human and economic resources for the mining companies.”

APC systems are very often seen as one of the main drivers needed to reach the dream of autonomous operations, Becerra explained. “In this context, it’s commonly heard in the media that AI is replacing APC systems,” Becerra said. “But this wrongly assumes that AI is already a synonym for fully autonomous operations. This kind of misrepresentation does not help, as such fully autonomous, continuous-process plants are still not that close to reality.

“However, there are many examples where new technologies and workflows can heavily enhance the level of information that is gathered and analyzed, transforming it into much better actionable insights, to take decisions faster than ever,” Becerra said. “This is what we call ‘intelligence augmentation’ and can clearly assist and elevate the performance of either existing APC systems or human-based control.

Becerra pointed to three main areas where APC could benefit from AI: Cognitive augmentation. The ability to gather, analyze and combine various data streams in real time can bring relatively quick benefits from operational and safety perspectives. One example would be building new virtual sensors to replace unreliable or unavailable signals, particularly when the instrumentation is placed in risky areas or is often out of service. Smart controllers. In certain contexts, controllers can be enhanced and complemented by virtual models of machinery or processes, known as digital twins. If the digital twins are done well, they can be used to find the controller’s optimum parameters, which leads to more stable processes, achieves higher production and quality levels, or decreases the amount of energy or water used. Dynamic adaptiveness. Many processes are by nature nonlinear and time-varying: this means that actions that were optimal to achieve specific goals yesterday (or even an hour ago) may be suboptimal or even inefficient now. Varying grades in the ore feed would be a classic example.

The new technologies and collaboration possibilities introduced in PXP V8.5 will bring multiple options for cognitive augmentation and smart controllers, Becerra explained. “For example, we are facilitating techniques to apply neural networks to build smart predictors, replace faulty or missing instrumentation, or forecast the future of certain signals in the form of time series,” Becerra said. “This creates the option to connect those new synthetic measurements with controllers, which will thereby react earlier, preventing undesired situations from happening or sticking constantly to optimal setpoints.

Becerra said two main barriers limit adopters from realizing the benefits. First, these new capabilities not only involve new technologies but also require new procedures, work flows and skillsets. It is therefore important to understand that multidisciplinary views and cross-functional collaboration are more crucial than ever. Process specialists (domain experts), automation and process engineers should open their arms to and work closely with data scientists, data engineers and industrial AI experts to explore potential new solutions to specific process problems. This human and social aspect is commonly overlooked but, in reality, working as a strong team of people with complementary skills is a key element to success.

The second aspect relates to the “hype cycle,” especially with emerging technologies and trends in the industrial landscape. “We hear bold promises from marketing materials or sales presentations sometimes inherited from other sectors where maturity levels and/or conditions are far from similar. This can make it very difficult for a non-technical audience to discern hype from what is technically viable and commercially profitable for their specific business needs. This overinflation of expectations, combined with low resistance to failure, leads to huge doses of frustration and early dropping of the investment, even before the learnings are incorporated into a new iteration or before a good productivity level is reached.

Getting Started
Applying AI and machine learning to the entire plant all at once could be time consuming and would not likely produce the intended results. For mining companies interested in getting started, Becerra suggested starting with a segment of the flowsheet, such as the grinding circuit because of its non-dynamic characteristics. In some cases, he said he has seen significant results, production increases of 6% and decreases of energy consumption of as much as 7%-8%.

Using machine learning with the grinding circuit, process engineers could use predictive modeling and time series forecasting and determine the influence of setpoints. Predictive modeling allows the process engineer to develop models to use in parallel with existing instrumentation. Plants measure P80 frequently and the PXP DataBook could add the capability of using six to eight months of data to predict P80.

With time series forecasting, process control engineers would have the ability to predict how the plant will operate in the next five to 10 minutes. Becerra likened it to weather forecasting. If nothing changes, according to the historical data, the mill weight will increase and they need to take action to keep it from overfilling in the next 10 minutes. This is not to be confused with predictive analytics that estimates a motor failure in three months, as an example. Time series forecasting is very short term. Machine learning can answer questions surrounding setpoints and the optimal parameters of operations. Questions such as: What mill feed and speed would achieve the lowest energy consumption? From those answers, process engineers can compare production as a function of energy consumption.

Looking toward the future, Becerra sees even more improvements ahead for PXP. With FLSmidth’s acquisition of KnowledgeScape in October, it now has access to applications related to thickeners and flotation. The company will soon integrate those into the platform.

Transform or Be Marginalized
What will the future of mining look like? Why does the industry need to change? How can mining embrace innovation? These were just some of the questions pondered when a panel of mining industry experts came together to discuss the need for change in the mining industry from an operator, solution, technology and investor perspective at’s recent “Inspirations20.” In the keynote session, Founder and CEO Sam G. Bose was joined by Damien Caby, senior vice oresident, Oilfield & Mining Solutions, BASF; Ippei Akiyoshi, CVC Mineral Resources Group, Mitsubishi; and Cleve Lightfoot, head of innovation, BHP. Central to this discussion was the notion that mining technology and processes have become outdated. “We’re dealing with technology and processes that were developed 100 years ago that aren’t apt for what comes next,” explained Lightfoot. “What we’ve done in between is we’ve made them bigger, better and more efficient and we’ve done a good job at that. What we haven’t focused on is how the context in which we operate has changed and it’s changing more and more rapidly.”

When designing new technologies for the industry, Bose explained that demand for sustainability solutions has become a top priority. Caby added that from an operational perspective, the demand for ethically sourced and environmentally conscious material is also heavily influencing the direction of the mining industry. “We see that, both as a supplier to the mining sector and a supplier to the battery industry, there are tremendous changes in the nature, the quality of metals that are going to be required for the world, which in itself is a major driver for change,” he explained. “I just don’t think we’re going to provide the nickel, the copper and cobalt for e-mobility the same way that 20 years ago the industry was able to react to the surge in demand for rare earths, and I think overall that’s the biggest driver of change.”

With the pressure to innovate or perish, panelists conceded that it may appear difficult for mining companies to know where to start, but prioritizing opportunities was key. “We’re great at building big infrastructure projects, but we need to learn how to figure out how we take little bits and transform through little bits. So, accept the ambiguity, and the risk associated with these things through bites we can actually manage,” Lightfoot said. Caby suggested that for mining operators unsure of how to start their innovation journey, the most pragmatic option may be to start with unlocking the value in their operational data. “In our experience, it really starts with understanding of the situation of the mine and really diving into the situation of what’s the challenge, and that’s all really data based,” Caby said.

“You can have much easier access to information, you can measure the impact of the improvements to make and correct the results for changes in feed and operating conditions, you can make simulations that really help address risk concerns. If you can run scenarios, if you can put boundaries, if you can try and see what can happen worse case on the computer compared to trying it real life, you really have a powerful tool to enable innovation.”

Watch the full recording of the session at https://inspirations.

As featured in Womp 2021 Vol 01 -