Most production loss and safety-related issues start with a small incident that escalates. How can we smartly use the data to mitigate small incidents and avoid disturbances? The goal is to operate as close to the theoretically possible limit, while at the same time being able to handle incidents.
How can we use the available production data in a new and smart way to avoid disturbances leading to loss of production?
This figure above shows how production can vary on a plant. Ideally, we want to reduce the time spent in the yellow regions. The control room operator is the one that has to deal with the plant disturbances when the automatic control system cannot control the plant. This is always the case, even if the level of automation is increased. In the foreseeable future, a human operator will still be needed.
Digitalization applied to existing data can provide valuable insights through big data analytics and visualization. We need to convert the data to insights that can assist the control room operator in the decision process, in order to be able to handle incidents better and increase situational awareness. The control room operator sees the process of the plant through the graphical user interface of the control system, which is assisted by alarms that provide symptoms of something being wrong. One of the major challenges in the control room is the complexity and size of the process and the technical “support” systems.
Will new technology improve production efficiency?
New technology can provide valuable insights, but new technology will have to be introduced cautiously. If we simply add new technical systems, we will be overloading the control room operator. The benefits of adding new technology in the right way could be to improve insight and understanding of the information that the operator receives when an incident occurs, which will help them to reduce downtime and risk.
Artificial intelligence and Big Data
The development of artificial intelligence has been proven useful within different industrial applications. Big data analysis techniques, using machine learning, and deep learning on existing data sets, can help provide insights into the process system, supporting the operation. With big data analysis, we are now able to analyze, recognize patterns, and learn from old data sets, which has not been manageable previously. However, we need to consider how the results from the analysis are validated and presented to control room operators.
How can we trust the results and how can we expect a human operator to believe in the results?
The way we utilize AI today falls into the category “Narrow AI”, meaning that AI is useful and may outperform humans, but only when applied to very specific tasks.
"A simple way to explain this is that in order to create a powerful program (deep learning or neural network) capable of recognizing cats, the AI algorithm must be provided between 100,000 and 1 million images of cats. In comparison, a human baby is able to recognize a cat after having seen just two or three images," - CEO of Total, Patrick Pouyanné (LinkedIn article March 4, 2019). The illustration below shows how easily we can recognize which images are of dogs and which are of bagels, but artificial intelligence will struggle with finding the difference with these few images.
Image from twitter @teenybicsuit
AI without human interference is far away from being able to successfully deal with general complex situations. AI will have to be used carefully in order to avoid negative side effects of following wrong objective function or extrapolating from too few samples. The negative side effects will be making worse decisions as the relationships presented are not valid, nor accurate. For further reading on the topic: Concrete Problems in AI Safety (https://arxiv.org/abs/1606.06565).
Data visualization techniques
Existing data may be visualized using new techniques in order to get a better understanding of situations quicker. The real-world environment is mixed with computer objects. A simple example is the Head-Up Display in new cars, where key data is projected onto the car windscreen.
Similarly, smart glasses with a projection of key data can be used to visualize real objects, for instance, fluid temperature inside a pipe for field operators. Key data or the result of analysis and algorithms, may be used with CCTV camera feeds for an enhanced understanding. Combining 3D models with images and real-time data can be used for situations where the operator may choose to see through a wall in order to gather more information.
The concept of digital twins is building digital replicas of real-world objects. The twin consists of data-driven analytical algorithms that are connected to available real-time and historical data. The digital twin may be used for organizing data and making documentation available at an entry-level. Additional use will include comparing results of algorithms with real-time observed data, identifying deviations, and predicting and preventing future incidents. An ideal twin can combine explicit first-order physical models with big data deep learning and ideally tacit domain knowledge, for improved decision-making.
Increase insight and not alarms
The technology development within digitalization has given us more tools to use on existing data for handling plant disturbances. We do not have to wait for additional data to become available through the development of IoT (Internet of Things = connecting all different things to the internet) and more, with data sources. The development of IoT will add further value to the next steps. Caution needs to be taken when introducing new tools in the control room. This is with respect to the quality of the decision support and the way it is presented. Adding one more screen with more alarms will not help the control room operator to make a better decision. It may even add complexity and confusion to the situation.
Click below to learn how we can use the next generation of digital twins to increase insights resulting in increased plant performance and safer operation: