“The goal of the data science process in the industry is real-time actionable insights from data-driven models in the operational environment,” said Dr Mark Tibbetts, data scientist at Arundo Analytics. The industries he was referring to are oil and gas, mining, power, utilities and maritime.
Arundo describes itself as a company that bridges the gap between data science teams and operations people, with the goal of “smarter operations through industrial analytics.”
With “data becoming available pretty much at the touch of a button,” according to Tibbetts, thanks to near ubiquitous sensor connectivity, the decreasing cost of data storage and computation, and migration to the cloud, heavy industry can make use of data science in many ways.
Those eight areas are; benchmarking and KPIs, predictive maintenance, production and consumption forecasting, efficiency and optimisation, health and safety, logistics planning, fraud detection and automation.
“Predictive maintenance is a very hot topic. You can predict a piece of equipment is going to fail before it happens."
“Predictive maintenance is a very hot topic. You can use the data to predict that a piece of equipment is going to fail before it happens. Then you can reoptimize maintenance schedules to avoid unplanned downtime,” he said.
He explained that a single manufacturer has the opportunity, in this case, to build a data science model from the analytics of a particular piece of equipment and then sell those analytical insights to the asset owners.
A challenge to collaboratively creating a data science model for one particular piece of equipment of a specific model and make is that out in the marketplace it will have different asset owners. The data models developed by the owners is likely to be siloed. “There’s commercial implications and competition considerations to take into account as to whether asset owner A is going to share their results with asset owner B,” said Tibbetts.
When Arundo is approached in to work with a company, Tibbetts and his team try to understand the right use case before they start doing the analytics.
“We want to identify the decision we’re going to influence – who makes it and how data can improve it."
“We want to identify the decision we’re going to influence – who makes it and how data can improve it. Then we need to identify the potential value or business impact of providing a perfect insight to support that decision,” he said. After that they need to understand what is the data that supports that decision, how much there is and how easily accessible it is.
Tibbetts said use cases are usually either very feasible or have high business value. However, in the beginning, if a choice has to be made, he would advise to opt for feasibility.
“Start with the more feasible use cases that are less value and get them in place. Some of the things that you thought were infeasible become more feasible as you understand the data,” he said.
When it comes to choosing a methodology, Tibbetts said, there’s a balance to be struck between performance and scalability. “You might be able to take a single piece of equipment and come up with a perfect model that has 99.9% accuracy of the insight you want to gain from it, but that method is not easily scalable,” he said.
A key lesson that he and his team have learnt is that everyone in the company has to be on board with data science, from the executives downwards, for industrial IoT to be a success. He said that the company’s decision-makers, executives and management who fund the data science initiatives and get the proofs of concept, should also foster a cross company culture where data scientists can flourish.
“Data science is going to be generating models which are providing insights and you need to communicate what those data-driven insights are."
He pointed out the importance of getting the people who do the operations on board. “Data science is going to be generating models which are providing insights and you need to communicate what those data-driven insights are and the context of those insights to the operators.” He said, conversely, the operators who have the subject expertise and understand the context behind the data, have to connect with the data science teams, so there must be a two-way flow of communication to fully reap the benefits of data science.
Tibbetts said that, as in most sectors, companies in heavy industry fall on different points along the maturity scale for internet of things and so there is a gap in the readiness of different companies when it comes to applying data science.
Those companies at the lower end of the scale are still determining how data science models are better than the incumbent engineering or physics-based models. They haven’t yet bought in to the concept and need to be convinced.
In contrast, the most mature companies are determining how to manage model reliability, uptime, version control, and accuracy across hundreds or more enterprise-scale models. They have a data science team in place that are driving insights and taking those insights to the operations people.
Tibbetts said that many companies are somewhere in between; they have done proofs of concept, they’ve accepted this is an area which is going to add impact and they are starting to think about how it can fit into an operational environment, how it can be scaled, how it can be rolled out across the company’s assets and how they can rapidly deploy these data science models across a variety of use cases.
With all this in mind, Tibbetts said it is important to manage expectations about what can be achieved with data science and in what time frame, to the executives of companies that are less mature. “I’ve gone into a lot of companies who have read Business Week about what AI and machine learning can do. I have to say, 'I’m not going to be able to give you an automated rig in six months',” he said.
Mark Tibbetts was speaking at the Data Science Institute at Imperial College London.