Invalid or missing GPS data can occur for various reasons including faulty equipment or loss of machine network connection at site. GPS coordinates are a primary source of information for the Digital Stockpile model and inaccuracies in the model will occur in event of frequent missing GPS information. Rectifying the root cause/s of the issue/s is the best way to handle erroneous raw data, however, the impact of the invalid/missing data is alleviated in the interim by GPS filling (neural network fill approach). GPS filling aims to improve the raw data and the resulting model.

To identify invalid GPS events that were meant for a stockpile, location labels need to be used as a secondary information source. Location labels are strings indicating the intended location of a dump or load. Conditioned exact label comparison is used for identification. Consider a theoretical stockpile, Moon A. The most commonly reported location labels on the stockpile are “A Moon”, “Moon A (0)” and “AMoon”. The labels can be conditioned (made lower case and removed special characters) to "amoon", "moona0" and "amoon". We uniquely have the conditioned identifier labels "amoon" and "moona0". If a new event occurs with the location label "A-MOON", the label would be conditioned to "amoon". Thus, we would identify it as belonging to stockpile Moon A.

Dump or load sequences on stockpiles tend to occur in patterns. Consider the anonymized x and y coordinates for a dump dataset plotted blow. Clear trends in both the x and y coordinates are observable.

The above data set is missing approximately 5 % of dump events. Missing events are indicated by red lines in the plot below. When an event has invalid GPS data and is assigned to a stockpile, previous event locations can be used to estimate its location. We can essentially exploit the patterned behavior of stockpile events to fill in missing coordinates.

A naive approach is to fill missing events with the last valid x and y coordinates. This approach can be sensitive to outliers and does not take advantage of the trended or patterned behavior of stockpile events. For the example data set, previous value fill results in the plot below.

A more robust method than previous value filling is to fill missing events coordinates with moving averages over previous x and y coordinates. This takes basic trends into account and is less sensitive to outliers. The moving average variant simplest from an implementation point of view is the exponential moving average. An event window of 10 events prior to the missing event was selected for the moving average. For the example data set, exponential moving average filling results are plotted below.

The previous value fill and exponential moving average fill methods are simple to implement, but cannot take advantage of more complex dumping and loading patterns. The GPS filling problem can be approached as a supervised learning problem. A neural network can be trained on historical data without missing GPS values. Training data can include spatial data, as well as material property information. This approach is expected to perform best and could identify more complex spatial patterns (as well as other correlations) within stockpile events. Using the neural network filling approach (window of 10 previous events) for the example data set, we obtain the plot below.