• Enhancement of algorithm for detection of gold strip circuit vessel sensor errors

      de Melo, Eduardo Pimenta; Ganguli, Rajive; Ghosh, Tathagata; Arya, Sampurna (2019-08)
      Sensors are used to understand the condition and flow of mineral processes. Having accurate and precise information is fundamental for proper operation. Even small errors are relevant to cost when considering the operational span of a mine. Finding small errors is hard; few algorithms can detect them and fewer still, when considering errors on the scale of 2% in magnitude. Some tools have recently been developed using data mining techniques for detecting small errors. Rambabu Pothina (2017) created an algorithm for detecting small errors in strip vessel temperature sensors in the carbon stripping circuit in Pogo mine. The algorithm performed well and was able to detect small magnitude errors without disrupting the industrial process. This thesis improves the understanding of the performance of the algorithm, while also making some minor changes. First, a statistical analysis of the results of the algorithm on baseline data revealed an inherent difference in how the carbon strip process was run with respect to the two strip vessels. This discovery provided insight into the algorithm, and how its performance depended on process characteristics. Second, the error detection algorithm was tested under scenarios different from Pothina (2017). Three separate types of errors were artificially added to real data: a) a fixed 2% ("fixed" error increase) b) a fixed 2% decrease ("fixed" error decrease) and c) an error with a mean value of 2% of magnitude ("noisy" error). Additionally, error was added to temperature data from each strip vessel (rather than just one), though only one at a time. The algorithm was tested under each scenario for each of the four years, 2015, 2016, 2017 and 2018. The time to detect errors ranged from 19 to 73 days. The time to detect was very high (53 to 73 days) in 2017 since there were large data gaps that year. In general, time to detect was about 30 days. The performance under noisy error were not that far below fixed error scenario. The algorithm took 10% more time to detect errors under noisy error scenario compared to fixed error scenario. On average, the algorithm detected an error after 25 cycles, regardless of the time span this represents. This is consistent in years with plentiful data, such as 2015, as well as years with less data, 2017 and 2018. In years with data gaps, 25 cycles represent a longer time period. Seeded errors that decreased vessel temperature have very similar results to its equivalent increase, i.e. the decrease in 2% of S2 has results similar to the increase of 2% in S1 and vice versa. In conclusion, these additional testing and analysis helped develop a more comprehensive understanding of the behavior of the data and the algorithms. These results validate and strengthen the findings of Pothina (2017).