Abstract

Intelligent signal processing methods play an important role in identifying 'novelty' in machine behaviour, thereby facilitating greater operational availability and early detection of emerging faults. Complex industrial systems often capture a high volume of data, with high dimensionality, and therefore clustering techniques have emerged as primary candidates for machine fault detection and classification due to their pattern-matching and data compression capabilities. To address well-known 'sensitivity' problems associated with classical clustering techniques, however, a real-coded Genetic algorithm (GA) is now considered that is integrated into a traditional K-means clustering methodology. The resultant hybridisation provides an enhanced search algorithm by incorporating the local search capabilities rendered by hill-climbing optimisation and the global search ability provided by a GA. Performance of the proposed algorithm is compared with other clustering techniques using two artificial data sets, and a benchmark iris classification problem. Results show that, in all cases, the proposed algorithm outperforms its counterparts in terms of global search capability and robustness. Moreover, the scalability of the proposed algorithm to higher-dimensional problems featuring a greater number of data points, is also considered, and validated using an application to provide a reduced data-representation of measurements taken from industrial gas turbine units, and for fault detection based on rundown vibration signatures.