Within a few years, maintenance engineers may be using an "intelligent" piece of software to predict mechanical failures in operating longwall mines, before they happen. The ambitious idea is the subject of a PhD research project being undertaken by postgraduate student Daniel Bongers.
Bongers is a CMTE scholar at the University of Queensland’s Department of Mechanical Engineering and his project is funded by ACARP (Australian Coal Association Research Program) and CMTE (CRC for Mining Technology and Equipment).
Loosely dubbed, the Decision Support System, the software will provide underground operators with suggestions rather than directives and is very different from comparable rule-based expert systems in one important way: the software constantly updates itself. This is because it is based on the principles of a neural net - software that can learn.
The software would theoretically plug into an existing Citect framework and be capable of feeding into data already collected by Citect systems. This would include information from sensors attached to equipment, including the shearer, AFC, BSL, crushers, and, in some cases, the longwall tail-end. Because it is monitored by a separate system, data from roof supports will not form part of this system.
An example of the way the software could be used is with shearer water pump electrical trips, which happen up to seven times a shift in some mines. Bongers said these pumps usually tripped for two reasons - a false alarm or clogging of the filter.
"The benefit of being able to predict the trip is you could classify it as a blocked filter or false alarm. Then, instead of wasting time figuring out what it was, the operator could just reset it and production could resume," Bongers said. "If it was a blocked filter and operators got warning before it tripped they could clear it before it tripped."
Because he was aware that longer-term implementation at operating mine sites could encounter operator resistance, Bongers spent time at various sites gauging the response from operators.
"Their main concern was that the system didn't come up with problems every two minutes that weren't really a problem because it was too sensitive. Secondly, they would want more information than 'this general failure is about to happen'. They want to know this particular failure is about to happen because the motor current is too high, compared with, say, the bearing temperature.
"Experienced operators would then be able to say, "Oh that's because we’re going up a hill at the moment', and choose to ignore it. I think we could overcome their fears if we could give them enough information to make up their own minds."
To date, Bongers has collected ten months of real data from the Dartbrook mine in the New South Wales Hunter Valley.
"We will use 30-50% of this data to train the system and get it running. Then we will use the rest of the data to see if it worked with data it’s never seen before," he said.
Bongers admits, however, that quantifying how many types of failures to predict is a difficult ask because of the wide variances between mines, their equipment, and their operating regimes. But because the system learns, this is less of a problem in comparison with rule-based expert systems that some OEMs have developed to aid with maintenance.
Hal Gurgenci, deputy CEO and program leader reliability and maintenance with the CRC for Mining technology and Equipment (CMTE) and the Department of Mechanical Engineering, University of Queensland, said rule-based expert systems work by assuming, in a fixed pattern, that a certain event is followed by another event (or various possibilities), and then another.
"To cover everything non-trivial the system rule base would have to be so large it would be impossible to debug and make it operational. There is nothing wrong with the expert approach - with the caveat that you have to find the expert, interrogate the expert to extract all those conditions, and then code them into the computer," Gurgenci said.
Another problem with expert systems is, who is the expert? Indeed, there are so many variables involved in predicting mechanical failure that a definitive predictive tool is probably not entirely realisable, as Gurgenci readily concedes.
"We will be content if the software can indicate something is happening when it is happening, in real time," he said.
Bongers is hoping to complete his PhD in a year's time. Thereafter, if successful the project will proceed to commercialisation. More information about this project and other similar projects can be found at the CMTE web site www.cmte.org.au.