Investigating the effects of varying cluster numbers on anomalies detected in mining machines

摘要

Anomaly detection is very important for the mining industry. If anomalies in mining equipment can be correctly detected to predict machine breakdown, mining companies will be able to reduce the cost of maintaining their machines. However, anomaly detection in this context is quite difficult considering the large volume of sensor data involved and unlabelled nature of the data. Clustering techniques have therefore been applied to analyse this problem, by dividing the data into normal and abnormal clusters. In this paper, we investigate the influence of using different numbers of clusters in clustering models, which include k-means, fuzzy c-means and the self-organising map, to obtain useful data patterns and classify the data into normal and abnormal types. Our aim here is to reduce the trigger of false alarm in the anomaly detection process. The data used in this study is based on real-world grease cycle data from a mining company in Australia. Our experimental results show that with more clusters, the number of anomalies detected tends to decrease for the clustering models considered. This means false alarms can be reduced by increasing the number of clusters used.