ABSTRACT: The paper discusses potential use of data mining techniques in mining. It reviews the basic techniques and methods of data mining and proceeds to identify possible mining applications of this methodology. In particular the paper proposes use of data mining to develop predictive capacity related to condition and performance of mining equipment. Other possible uses of data mining include optimization of mine performance as well as equipment operator training.
Modern mine control systems and mine equipment are highly computerized. One result of this situation is that large volumes of data are collected that define mine performance and equipment condition. Some of this data is processed in real time to provide information that allows for optimization of mine performance. Examples are the fleet dispatch systems that develop equipment assignments, best matched to the stated objectives of the mining operation and based on real-time processing of data that defines equipment status and location.
Most of the collected data, however, is used for reporting and post-mortem analysis of mine performance, for equipment failure analysis and for prevention of its catastrophic failures only. An example are the vital signs monitoring systems installed on larger pieces of mining equipment. These systems collect data generated by a variety of sensors and store it to facilitate easy failure diagnostics. In addition these systems have a capability to warn the operator of impending failure or to conduct orderly equipment shut-down if an emergency situation occurs.
Availability of huge databases and spreading computerization has led to large strides in data processing capabilities and techniques. Variety of powerful data processing methods have been developed over the last years that facilitate rapid processing of voluminous data for extraction of user friendly information. One of such methods is data mining. Originally developed by intelligence community to look for information in huge communication databases, data mining has since found a range of commercial and scientific applications. Nowadays it is widely used by retail industry to analyze sales, direct promotion and marketing efforts, by cellular telephone companies to assure client retention, by scientists to search for information in large databases created by Hubble space telescope, and in many other applications.
This paper briefly reviews data mining and the related techniques, and proposes their use for discovery of knowledge in data acquired by a variety of data acquisition systems used in today’s mines. In particular the paper suggests that data mining can be used to develop predictive capacity related to equipment condition and its performance. Data mining offers a potential for further, significant improvement of mine performance.
Data mining is an iterative process that involves setting the objectives of the search, selecting and cleaning input data, transforming it, running a mining function and interpreting the results. The schematic in fig.1, adopted from IBM (International Business Machines, 2000), presents these tasks graphically.
The selection of data to be analyzed may involve integration of data from various sources and often requires their formatting to fit the format acceptable to the data mining software. In a mining situation where the objective may be optimization of Komatsu truck performance, data on load carried, on cycle times, and on truck component performance may be needed, acquired in different formats from engine monitoring system (say Cummins engine monitoring system), from truck dispatch system (say Modular Mining’s Dispatch), and from an on-board weigh
Figure 1. The data mining process
measuring system provided by a third party. Major problem may be faced with making data formats compatible with each other and with that of data mining software to be used.
The next step, transforming the data or its pre-processing may involve filtration, discretization, data joining and similar actions. It allows organization of the data so that it may be mined efficiently. In the case of Komatsu truck mentioned above the data joining would be a major task, as would its discretization and filtration.
Mining data is done using one or more of data mining techniques briefly discussed below. It needs to be noted that data mining did not originally relate to mining. It is a general-purpose data processing method that permits discovery of information that may exist in various databases.
Interpreting the results is the last and a very important step of data mining. Usually various visualization tools are used in the process, which allow for easy viewing of the information and identification of information discovered during the data mining process.
Data mining techniques
A number of techniques are used in data mining, each with its own interesting applications. Several textbooks summarize and describe these techniques (Berson and Smith, 1997, Westphal and Blaxton, 1998, Weiss and Indurkhya, 1998, others). As an example. Berson and Smith (1997) classifies data mining techniques as follows.
The decision trees are predictive models that an be viewed as a tree, with tree branches representing a classification question and the leaves representing partitions of the data set with their classification. The prediction is made on the basis of a series of sequential decisions. Thus in case of mining trucks the decision tree could be used to identify which trucks are most likely to fail, and when, based on such questions as: what is the truck make, how old is it, how long it has operated, what is its past repair history, who was its operator and the like. A decision tree model can be confirmed or modified by hand and it can be directed based on the expertise of the person constructing it.
The decision tree models are best used for exploration of the data sets and that of the problem at hand. It is done by looking at the predictors and values that are chosen for each split of the tree. They can also be used for data pre-processing for other prediction algorithms. An example of such application is shown in the companion paper (Golosinski et al, 2001).
Neural networks are computer implementations of sophisticated pattern detection and machine learning algorithms used to build predictive models from large historical databases. They allow for construction of highly accurate predictive models that serve to solve a large number of different problems. The main problem with neural modeling is lack of clarity, the price often paid for their complexity and high accuracy. To overcome this problem, various visualization techniques are used in conjunction with neural models to help explain and control the model.
The primary application of neural models in data mining is clustering, the technique that is used to segment a database into clusters, or sub-sets, based on a set of predetermined attributes. The ability of neural models to perform accurate numerical predictions led to variety of applications, including predictions of the stock markets behavior. As related to a mining truck, neural clustering may be used to define and quantify the relations between various data streams collected on this truck, following by clustering of these streams into mutually dependent groups. Thus, for example, the factors that have an impact on cycle time of the truck can be defined and quantified.
2.3Nearest neighbor and clustering
Both these techniques are very intuitive and between the first used for data mining. Nearest neighbor prediction algorithms are convenient and simple predictive tools that allow for clear explanation of why a prediction was made. The predictions are based on behavior or properties of the “neighbor” data with the highest weight assigned to the data that is closest. Clustering is grouping, or “clustering’ together the data that has the same or similar attributes.
Both clustering and nearest neighbor techniques are between the easiest to use and have a variety of applications. Both are primarily used for prediction of new data rather than extraction of rules from an extensive databases. Using the mine truck example, these techniques appear to be most suited for prediction of when and how this truck will fail, a key piece of information for a mine operator.
Genetic algorithms refer to simulated evolutionary systems that dictate how populations should be formed, evaluated and modified. One of a variety of algorithms known as optimization techniques generic algorithms are in their infancy and more experience with them is required before a mine-related use can be proposed.
Rule induction is one of the most common forms of knowledge discovery in unsupervised learning systems. This technique is often used to “mine” databases, to discover information that is not obvious or readily available. The technique retrieves all potentially interesting data patterns in the database with the found rules being generally simple and easy to understand.
The rule induction can be used to make predictions, but its main use is for unsupervised learning to find rules that are not already known. In reference to the mining truck the rule induction may be used to define relations between various data streams collected on this truck. As an example a rule can be discovered that states: “if this truck is operated by operator x and it is Monday, the performance of the truck will be dismal”. Likewise a rule can be defined that states “if the truck engine overheats and strut pressures are within certain range, the truck is overloaded”.
This technique offers a great promise if applied to mining equipment operator training.
Use of statistics is by far the most common approach to data analysis and various statistical theories and calculations can be used to discover hidden patterns in the databases. These include, but are not limited to regression, curve fitting, principal component analysis, factor analysis and other.
As the statistics is one of the well established sciences and a huge volume of information on its application to pattern discovery is available, this data mining technique is not discussed further in this paper.
3MINING USES OF Data mining
The focus of data mining is to discover and define hidden patterns and trends. Once a pattern is defined it can be used in many ways, such as a training input into a neural network or encoded as a rule into an expert system. Traditional applications of data mining include those for monitoring medical bill fraud, marketing with coupons, monitoring credit card transactions, and the like (Westphal and Blaxton, 1998).
The data mining is estimated to be a $20 billion industry today. In spite of this, to the best knowledge of the author, no attempt was made to use data mining techniques to address mining related problems so far.
Huge volumes of various data are collected on today’s mining equipment. As and example each large off-highway truck manufactured by Caterpillar is equipped with the so called VIMS (Vital Information Management System) system that has a capacity to collect, store and transmit information from over 150 sensors installed throughout the truck. With the sensor indication sampling rate of one per second, and truck operating 7,000 hrs per year, over 3,780 MB of data can be collected for each truck during one year of its operation.
While some of this data is used to generate information describing truck performance and condition, most of the collected data remains unused and is not analyzed. Very little of it, if any at all, is used to forecast truck condition or performance into the future. Instead the whole data analysis effort directed on assessment of past performance. Use of data mining techniques for information discovery in this huge database appears to be one of the promising ways to improve performance of many mines.
Review of current industrial applications of data mining indicates that there are numerous opportunities for its use in mines. Three most obvious applications are (1) mining equipment condition monitoring and failure prediction, and (2) quantification of and prognostication the mining equipment performance (3) training of equipment operators.
This application offers the highest potential for successful application of data mining in mining. The approach judged most promising is to (1) find, define and quantify the relations between various indicators of equipment condition based on data mining of the data collected by relevant sensors, and (2) use the discovered relations to build predictive models that would permit prognosticating future equipment performance.
Data mining techniques of clustering and association appear to be the most promising in defining the relations and associations that may be of interest. On the other hand rule induction and polynomial regression, the latter not discussed here, may be the best techniques to develop the predictive capability.
In addition to equipment condition related data, variety of performance related data is available for each piece of mining equipment. This data is collected though fleet dispatch systems now used by a majority of surface mines and some underground mines. Alternatively, this data can be collected by on-board monitoring systems, an example being Caterpillar VIMS system discussed above. If installed on a mining truck the VIMS collects data on truck load size, truck speeds, and the like. It also calculates cycle times and other truck performance related data, and stores all for downloading or transmittal to mine databases.
Similar to equipment condition monitoring, discussed above, the database that contains equipment performance data can be mined for pattern discovery. Discovery of patterns which undoubtedly exist in this database may then permit construction of a model able to prognosticate performance of the mining equipment under a variety of scenarios. While this concept is somewhat similar to fleet simulation models that may be a part of the dispatch system it offers a number of added benefits. These include, but are not limited to, ability to set performance standards for future enforcement and to define the optimum operating parameters for various pieces of equipment.
As an extension of data mining use for mine performance improvement, hidden pattern and trend discovery can be used to design and implement more effective operator training program. Based on quantified patterns and trends the optimum operator responses to various operations conditions can be defined and communicated to the operator. This may include definition of optimum speed at a specific segment of a haulroad, definition of optimum load, definition of the optimum accelerating and braking patterns, and the like.
Modern mines generate huge quantities of data that describe and quantify condition and performance of mine equipment and of the mines themselves. Availability of this data creates a unique opportunity to improve performance of both.
Data mining, a set of techniques used to discover hidden relations and trends in large databases, is the likely tool that will permit this to realize this opportunity.
The most obvious mining applications of data mining are to prognosticating condition of mining equipment, to prognosticating its performance and to training of equipment operators.
Berson, A. and Smith, S.J. 1997. Data warehousing, data mining and OLAP. McGraw-Hill
Golosinski, T.S., Hu, Hui and Elias, R. 2001. Data mining VIMS for information on truck condition. APCOM 2001, Beijing, China.
International Business Machines Corp. 1999. Using the Intelligent Miner for data. Company publication.
Weiss, S.M. and Indurkhya, N. 1998. Predictive data mining. Morgan Kaufman Publishers, Inc.
Westphal, C. and Blaxton, T. 1998. Data mining solutions. John Wiley & Sons, Inc.