Building a truly predictive model is the holy grail for many miners. Predictive maintenance expert and DINGO director of product engineering, Colin Donnelly, tells how it can be done in an interview with Noel Dyson, editor of Australia’s Mining Monthly.
One of the keys to getting predictive maintenance to work is in the data, obviously. However, it is knowing what data to keep that can be the trick.
Donnelly said it was important to keep records of failure types. “In most cases people don’t record the failure reasons in any depth,” he said. “To be effective we need to link things to past behaviour. If you have enough well categorised data you can build a strong and robust model you have confidence in.”
As an example, Donnelly points to a case in Canada DINGO worked on. DINGO took 200 rebuild documents and recategorised them. “We had a few of our subject matter experts go through the documents and identify the primary cause of failure.” These experts assigned a standard rating for each of the parts in the engine.
“Once the information is properly categorised, we can let the machine learning models take over” Donnelly said. “It can compare the condition monitoring data with the failure data. From that we can create a remaining useful life and give a probability of failure.”
DINGO works with a wide range of mining equipment, and sensors on these machines report a wealth of useful information on how their engines and other systems are operating at any given time.
“From there we look at how that data relates to failures and can use machine learning to determine probabilities of failure,” Donnelly said.
“We also use predictive analytics to find things a person wouldn’t notice. We’re trying to find the very small deviations.”
Better data categorisation is a way towards building better models.
DINGO took part in a predictive analytics challenge last year that looked to develop a failure prediction model for fixed plant equipment. In that challenge, Donnelly said DINGO was up against some of the major players in the machine learning space.
He said the miner packaged the data up perfectly for all the participants.
“We had a very finite set of data that lined up specifically to the failures” Donnelly said. “It still took three months to get the model to 90% accuracy.”
In the end three companies got through the hurdles and DINGO was one of them.
So how to get the data into the right format? In Donnelly’s view, an Excel spreadsheet can be sufficient, provided the right information is accurately logged.
However, he warned that predictive analytics was not the sole answer.
“All predictive analytics will do is predict the future based on what it already knows,” Donnelly said. “It does not factor in intervention.
“I think everyone thinks computers will do everything for them. You need to have both the operating and
the failure data.”
Donnelly said DINGO was working to use Machine Learning and AI for practical applications that would lead to tangible outcomes.
The company comes at maintenance problems with 25+ years’ experience with mining equipment under its belt.
“We have people who have been assigning corrective actions to address specific problems and we’ve tracked these issues to resolution ” Donnelly said. “We have a database of actions people have taken and the impact it had on the machine. and we can predict outcomes with a high degree of confidence.”
Size does matter though.
Donnelly said those mines with small fleets would likely benefit from working with a company like DINGO, which has access to a massive data set.
“Big mines with bigger fleets may be able to make some progress using their own data, but DINGO has the advantage of well categorised data from equipment operating at mines spanning the globe.”
Noel Dyson │ Australia’s Mining Monthly │ January 2019