Data mining mine data: truck-shovel fleet management systems

CIM Bulletin, Vol. 2, No. 4, 2007

S. Dessureault, M. Yildirim, and M. Baker

The double effect of high commodity prices and improvements in information technology (IT) has resulted in more mining companies investing in IT. One of the traditional IT process control tools used in the mining industry is real time in-pit truck allocation control and monitoring systems, or Fleet Management Systems (FMS). These systems have been used to maximize the overall mine production by improving equipment utilization and reducing the production costs. Most modern operations monitor the in-pit operations using GPS and wireless communications that produce a massive amount of data. The data is used in real-time to calculate truck allocations using optimization algorithms and to generate reports, including production and machine health reports. With ongoing improvements in bandwidth, sensors, and computing power, modern systems are becoming more complex and generate increasingly more data. However, human analysis by simple regression or table generation in spreadsheets is becoming unwieldy because the data sets and number of variables is becoming too large. A data warehouse (DW) is used in many sectors as a key piece of IT infrastructure to collect and manage extremely large data sets. Once stored in the warehouse, one can undertake analyses such as reports, online analytical processes (OLAP), or more complex analyses such as data mining (DM). DM is a general term for applying algorithms, many of which pre-date the term data mining, to large data sets to ‘mine mountains of data to find nuggets of information.’ For example, packaged neural network algorithms within data mining software undertake the usual functions of a neural network, namely to estimate predictions then learn from the feedback to adjust the prediction model, which becomes increasingly more accurate as a result. Note, however, that in this application of DM, the neural network algorithms are applied to very large data sets. This industry-funded project had two primary foci: first, to analyze and contrast the performance of each commercial FMS; and second, to undertake knowledge discovery in the database by mining the historical data to discover patterns that may benefit decision support. Two different FMSs were used in the study, necessitating the need for a common data structure. A star-schema data model was designed and populated with the historical data from both FMSs. After preparing the fact and dimension tables, OLAP cubes are created to be able to undertake preliminary analyses. Several data accuracy and structural issues were identified. Filters were used to either eliminate faulty records or fix the errors. Once the data was properly structured and largely cleaned of errors, the data mining could begin. A data mining tool called a dependency network was used to analyze the importance of the different operational variables on particular performance measures. The figure illustrates the dependency network of these operational mine variables on ‘waiting time at the shovel’ (truck queue time) with some non-representative delays filtered out. This paper describes the application of modern data mining techniques on truck dispatching systems in a real open pit hard-rock copper mine. The long-term benefits of this work are to identify strengths and improvement opportunities in truck assignment algorithms and to establish the skill sets and IT infrastructure needed to undertake more complex data-driven technology, such as a truck dispatcher trainer.