The introduction of Artificial Intelligence (AI) to mining – and in particular to mineral processing – is proving crucial to optimisation and as part of the industry’s journey to remotely monitored and truly autonomous processing operations. There are a number of companies involved both large and small. IM recently spoke to one of the innovative smaller players in this space – Malmö, Sweden-based Sentian.AI and its CEO Martin Rugfelt, about the true potential of AI in mining concentrator plants when compared to manual optimisation or the use of advanced process control (APC) software. Most recently, Sentian.AI has become a member of SynerLeap, ABB’s innovation growth hub located in the heart of ABB’s Head Office in Västerås, Sweden. The purpose of SynerLeap is to create an ecosystem where ABB can utilise and enable partner technology companies to grow and expand globally in industrial automation, robotics, and energy. As a SynerLeap member, Sentian.AI will collaborate with ABB in two areas. Firstly, Sentian.AI will bring its cutting-edge AI-powered process optimisation solutions to industrial customers looking to reduce energy consumption, lower emissions, address a shortage of expert operators, and make production processes more efficient. Secondly, it will bring its AI-based planning optimiser, which supports the optimisation of logistics, inventory, and supply chains to the ecosystem.
Q Can you give some background into Sentian.AI as a company in terms of early development and specialising in industrial process optimisation and specifically mining?
Sentian.AI started after the core team exited another AI startup, after it was acquired by a large US technology company. Already then the founding team had decided to focus on the industrial side of AI. The reason was simple: industrial companies have big challenges that are suitable for AI, they have lots of data and AI solutions have a massive commercial impact when implemented. Our expertise was in optimisation and machine learning and we had already then worked on projects optimising the net present value of a mine, on predictive maintenance of mining equipment and equipment utilisation optimisation. Early on in Sentian’s development we also helped one of the mining OEMs with their AI strategy and after that mining was a natural focus area. One of the most interesting points is that the control systems are essential for efficient operations as they affect yield performance, energy costs, emissions, quality and throughput. We also saw that AI would be ideal for supporting operators to get more from the control systems and the process. It was therefore natural for us to focus on optimising mining processes and their control systems with AI.
Sentian.AI CEO Martin Rugfelt
Q How does your approach differ from “traditional” AI in industry?
A lot of the “traditional” AI systems we see are actually AI toolkits/platforms that are sold on the basis that the customer can create AI logic that they need to solve specific problems without having to understand the detailed data science. Unfortunately, without knowledge of the underlying AI and data science, creating AI for complex systems eg control processes is very difficult. As a result, many “traditional” AI systems are abandoned after purchase. Sentian has taken a different approach. Our SentianController is explicitly designed to optimise control of industrial processes, so when customers buy it, they already have the AI algorithms developed and tested for the complexity of control system optimisation. It is effectively a point solution that means you do not need a large data science team to build and run the AI solution.
Q Is your system based on unique/patented algorithms?
In the AI field, there are a multitude of different algorithmic approaches that have been developed. These algorithms perform very well for specific tasks and not so well for others. Sentian.AI has worked for many years to select the best algorithms and refine how these algorithms work to deliver a unique AI system that has been designed to control and optimise industrial processes. This is very complex and requires some of the latest technologies in AI to be able to achieve the necessary control. We have also developed a system that can uniquely be applied in stages as data quality improves and operator confidence increases – going from making recommendations to fully autonomous control at the speed our customers want. This allows customers to build confidence before committing to fully autonomous control. We have chosen not to patent our solution as it would have exposed the technology, however, we would argue it is very unique.
Q Can you give any examples of successes to date in specific areas of minerals or metals processing and the upside your were able to bring?
I think two projects we are working on are good examples of the applicability of AI in mining. I would like to highlight a flotation project and a crushing and grinding optimisation project. In the flotation project we are planning the implementation of the solution after a successful proof of concept during the spring of 2022. The target here is not only the increase of the recovery rate but also the stabilisation of the circuit performance. We are currently also working on optimising a crushing and grinding circuit. Here the primary goal is the increase of throughput.
Q When you say your system is able to self-adapt, can you explain what that means and how it is able to do this?
Self-adaptation can be achieved when you have both the right data and the right AI models. The AI creates a dynamics model that is made from both historical and “live” operational data. It can then choose the set of control parameters that deliver optimum performance towards a specific goal, eg maximum production for minimum energy usage. In comparison to traditional supervisory control systems it adapts to changes in the process. For example, if the process changes for some reason, leading to new data points, the AI incorporates these into its model, new predictions are made, and new control parameter settings are used. New goals can also be set, resulting in SentianController choosing the best control parameters to achieve those goals.
Q Is your ideal scenario to retrofit to an existing mineral processing operation or to be part of the plant from its start-up and commissioning?
Our ideal scenario is to retrofit our SentianController to existing mineral processing operations as there is historical data to train the AI. Also, you can more easily show the before and after plant performance. The solution of course can be added to a new build, but it takes time to build the necessary data in stable operation.
Q Is your system truly vendor agnostic and how does it achieve this?
As long as we can access the data we can implement the solution. We connect to existing data sources like sensors, DCS, SCADA, and Historian systems. If the data exists in the system, then SentianController can use it and optimise the process based on that data. We can also add new data eg if you add a camera system that delivers the flow rate or a variable speed apron feeder, that can be incorporated into the data set and used to control the process irrespective of who supplies parts of the system. Our solution uses simple, end-to-end secure, industry-standard web APIs, REST and gRPC and the solution can be implemented as a cloud solution or on premises.
Q How much historical data does Sentian AI need to “learn” a new process in mining?
That depends on the goals. For a fully autonomous solution there is a need for large amounts of high quality data. However you can also work with little data to achieve some less challenging goals than full automation such as providing estimates of how specific setpoint combinations will impact the process or proposing certain setpoints to operators. We suggest mining companies should take a stepwise approach where the goals are set according to the available data and then when new data sources are added to set higher goals over time. This means the benefits of AI can be gained much sooner than with large scale projects with lofty goals.
Q What problems exist with current process optimisation in mineral processing in terms of performance and functionality and how does Sentian help address these?
Today, optimisation of mineral processing is done either manually by the operators based on their knowledge or via APCs. Manually optimising process parameters can be a workable strategy with sufficiently experienced operators. The challenge is that plant performance can vary by operator experience or shift. It is also very challenging for any operator, regardless of experience, to see feedback loops between equipment and process stages, as the interactions are sometimes so complicated that a human simply can’t see them. An AI-based solution is based on data so it can model all the process dynamics, whether obvious or hidden. It also provides consistency across operators and shifts, so performance does not vary in the same way. In traditional systems such as APCs, control is usually based around an abstract process model. If the process changes or goals change, that model frequently needs manually updating by APC experts; otherwise, performance benefits degrade over time. The self-learning and adaptation built into AI systems such as SentianController means that the process can change, for example when a piece of equipment is replaced. SentianController updates its model and responses as it gets new data showing different control responses. In addition to adapting to changing conditions it also learns the optimum settings and continuously improves over time and has a smaller steady state error. It is also worth saying that we still see a crucial role for plant operators. Instead of controlling process parameters, their time is freed up to focus on identifying and fixing anomalies, identifying potential process bottlenecks, and ensuring the underlying systems are working correctly.
Q Is predictive maintenance an important aspect of what Sentian AI can achieve?
Sentian.AI offers predictive maintenance too, but not as a product. It is an area that is complex to productize. We offer consulting support in this area. Based on our experience in predictive maintenance it is important to customise the solution to the circumstances and pick the right AI algorithms carefully to get the relevant results. Depending on the quality and volume of the available data a combination of unsupervised, semi and supervised AI can be put together to create a solution that is effective. The attached graph showing set point changes indicates the fast reaction time and the stable production of SentianController while the multi loop controller is trailing far behind in adoption of the changes.
