“At Lihir, our three mills had been experiencing multiple overload events each year, resulting in significant downtime. Traditional engineering-based approaches had failed to adequately identify when overload events would occur. Over 360 million lines of data across 130 variables were collected, and by cross-referencing this big data, the Petra team was able to identify causal factors and develop machine-learning algorithms which could predict future outages,” he said.

The Lihir mill FORESTALL overload algorithms were developed using an engineered approach to data science. For example, in this case, engineering knowledge was used to create an additional 650 engineered signals from the original 143 raw signals for each SAG mill (about 800 signals per mill). Engineered signals in the algorithm development significantly increased the accuracy of the algorithms. 

In addition, PETRA principal and MD, Penny Stewart shared a few FORESTALL case studies during her ‘Silent Music: Mining Case Studies in Machine Learning’ presentation at Austmine 2017. She explained how the Lihir SAG mills are the symphony orchestras of the processing plant, producing hundreds of silent signals.

“People are very good at hearing when music is out of tune. But, when we see music as a series of signals, it is impossible to detect when something is wrong. Likewise, when we see machine signals, or silent music, it is very difficult to detect when something is wrong just by looking. We can think of machine learning as a mathematical ear listening to the silent music of machines,” she explained.

Just like when we hear a song over and over we get better at detecting when something is played out of tune, machine learning is able to distinguish normal operation by learning what is a normal machine signal, and what isn’t.

Click here for the full transcript of Sandeep Biswas’ presentation.