Automatic detection of sensor calibration errors in mining industry

Abstract

Sensor errors cost the mining industry millions of dollars in losses each year. Unlike gross errors, "calibration errors" are subtle, develop over time, and are difficult to identify. Economic losses start accumulating even when errors are small. Therefore, the aim of this research was to develop methods to identify calibration errors well before they become obvious. The goal in this research was to detect errors at a bias as low as 2% in magnitude. The innovative strategy developed relied on relationships between a variety of sensors to detect when a given sensor started to stray. Sensors in a carbon stripping circuit at a gold processing facility (Pogo Mine) in Alaska were chosen for the study. The results from the initial application of classical statistical methods like correlation, aggregation and principal component analysis (PCA), and the signal processing methods (FFT), to find bias (±10%) in "feed" sensor data from a semi-autogenous (SAG) grinding mill operation (Fort Knox mine, Alaska) were not promising due to the non-linear and non-stationary nature of the process characteristics. Therefore, those techniques were replaced with some innovative data mining techniques when the focus shifted to Pogo Mine, where the task was to detect calibration errors in strip vessel temperature sensors in the carbon stripping circuit. The new techniques used data from two strip vessel temperature sensors (S1 and S2), four heat exchanger related temperature sensors (H1 through H4), barren flow sensor (BARNFL) and a glycol flow sensor (GLYFL). These eight sensors were deemed to be part of the same process. To detect when the calibration of one of the strip vessel temperature sensors, S1, started to stray, tests were designed to detect changes in relationship between the eight temperature sensors. Data was filtered ("threshold") based on process characteristics prior to being used in tests. The tests combined basic concepts such as moving windows of time, ratios (ratio of one sensor data to data from a set of sensors), tracking of maximum values, etc. Error was triggered when certain rules were violated. A 2% error was randomly introduced into one of the two strip vessel temperature data streams to simulate calibration errors. Some tests were less effective than others at detecting the simulated errors. The tests that used GLYFL and BARNFL were not very effective. On the other hand, the tests that used total "Heat" of all the heat exchanger sensors were very effective. When the tests were administered together ("Combined test"), they have a high success rate (95%) in terms of True alarms, i.e., tests detecting bias after it is introduced. In those True alarms, for 75% of the cases, the introduction of the error was detected within 39.5 days. A -2% random error was detected with a similar success rate.