Time Series Analysis and Classification with State-Space Models for Industrial Processes and the Life Sciences

Abstract

In this thesis the use of state-space models for analysis and classification of time series data, gathered from industrial manufacturing processes and the life sciences, is investigated. To overcome hitherto unsolved problems in both application domains the temporal behavior of the data is captured using state-space models. Industrial laser welding processes are monitored with a high speed camera and the appearance of unusual events in the image sequences correlates with errors on the produced part. Thus, novel classification frameworks are developed to robustly detect these unusual events with a small false positive rate. For classifier learning, class labels are by default only available for the complete image sequence, since scanning the sequences for anomalies is expensive. The first framework combines appearance based features and state-space models for the unusual event detection in image sequences. For the first time, ideas adapted from face recognition are used for the automatic dimension reduction of images recorded from laser welding processes. The state-space model is trained incrementally and can learn from erroneous sequences without the need of manually labeling the position of the error event within sequences. %The limitation to weakly labeled data helps to reduce the labeling effort. In addition, a second framework for the object-based detection of sputter events in laser welding processes is developed. The framework successfully combines for the first time temporal change detection, object tracking and trajectory classification for the detection of weak sputter events. %This is the first time that object tracking is successfully applied to automatic sputter detection. For the application in the life sciences the improvement and further development of data analysis methods for Single Molecule Fluorescence Spectroscopy (SMFS) is considered. SMFS experiments allow to study biochemical processes on a single molecule basis. The single molecule is excited with a laser and the photons which are emitted thereon by fluorescence contain important information about conformational changes of the molecule. Advanced statistical analysis techniques are necessary to infer state changes of the molecule from changes in the photon emissions. By using state-space models, it is possible to extract information from recorded photon streams which would be lost with traditional analysis techniques.