Optimised meta-clustering approach for clustering Time Series Matrices

The prognostics (health state) of multiple components represented as time series data stored in vectors and matrices were processed and clustered more effectively and efficiently using the newly devised ‘Meta-Clustering’ approach. These time series data gathered from large applications and systems in diverse fields such as communication, medicine, data mining, audio, visual applications, and sensors. The reason time series data was used as the domain of this research is that meaningful information could be extracted regarding the characteristics of systems and components found in large applications. Also when it came to clustering, only time series data would allow us to group these data according to their life cycle, i.e. from the time which they were healthy until the time which they start to develop faults and ultimately fail. Therefore by proposing a technique that can better process extracted time series data would significantly cut down on space and time consumption which are both crucial factors in data mining. This approach will, as a result, improve the current state of the art pattern recognition algorithms such as K-NM as the clusters will be identified faster while consuming less space. The project also has application implications in the sense that by calculating the distance between the similar components faster while also consuming less space means that the prognostics of multiple components clustered can be realised and understood more efficiently. This was achieved by using the Meta-Clustering approach to process and cluster the time series data by first extracting and storing the time series data as a two-dimensional matrix. Then implementing an enhance K-NM clustering algorithm based on the notion of Meta-Clustering and using the Euclidean distance tool to measure the similarity between the different set of failure patterns in space. This approach would initially classify and organise each component within its own refined individual cluster. This would provide the most relevant set of failure patterns that show the highest level of similarity and would also get rid of any unnecessary data that adds no value towards better understating the failure/health state of the component. Then during the second stage, once these clusters were effectively obtained, the following inner clusters initially formed are thereby grouped into one general cluster that now represents the prognostics of all the processed components. The approach was tested on multivariate time series data extracted from IGBT components within Matlab and the results achieved from this experiment showed that the optimised Meta-Clustering approach proposed does indeed consume less time and space to cluster the prognostics of IGBT components as compared to existing data mining techniques.