I worked as an intern at Siemens Corporate Research, Princeton from April 2010-Feb 2011. My task, broadly, was to design and implement algorithms/heuristics for mining patterns from logs generated during the operation of various medical equipment (Computed Tomography machines, X-ray machines) at Siemens Healthcare. I have been involved in various projects, some of which are the following:

Designing Efficient Heuristics for Detecting Bursts from Event Logs (Sept 2010-March 2011, with Fabian Moerchen and Ioannis Akrotirianakis): While mining patterns from timseries data generated by healthcare equipment, we observed occasional bursty patterns in the occurrence of the events, i.e., there are bursty windows in time where at least k events occur within an interval of length at most t. Our goal was to search for a critical threshold pair (k*,t*) such that the number of bursty windows meeting this threshold pair is significantly higher than that for any other combination of threshold values. An exhaustive search over the entire possible range of k-values and t-values is impractical, hence we worked on the design of efficient heuristics for the search. The findings of this research got accepted for publication in DaWaK 2011. We also filed a patent application (# US 61/467,529) for it.

Zhu and Shasha [1] addressed the problem of elastic burst detection. While they acknowledged that finding the thresholds for labelling a window as "bursty" is part of the problem, they associated a unique threshold with each possible window size. We make no such assumption, and formulate the problem as a search problem in a two-dimensional space. Our problem formulation, as well as approach to the solution, are significantly different from [2] also, which introduced a hierarchical model to address the problem of burst detection, since a long burst might contain several smaller bursts within itself.

Analysis of CT Machine Logs for Predictive Maintenance (Summer 2010, with Dmitriy Fradkin and Fabian Moerchen): The X-ray tube is one of the most important and expensive components of the Computed Tomography (CT) machines. The tubes have to be replaced and serviced regularly for routine maintenance. Predictive maintenance of the CT machines becomes easier if it can be predicted when the tube should be replaced next time. Since we only knew when in the past tube replacements had taken place, we had to rely on model-based anomaly detection techniques. We trained a multidimensional Gaussian Mixture Model using data (on temperature, current, voltage etc) from time-windows much before the replacement, and then derived the likelihood values of the data points based on these models as the dates approached the replacement date. For a significant number of machines, we noticed a steady decline in the likelihood values as we approached the replacement dates.

Design and Implementation of a Text Parser for Processing Maintenance History of CT Machines (Summer 2010, with Dmitriy Fradkin and Fabian Moerchen): We worked with logs obtained from Computed Tomography (CT) machines where each machine had a number of photomutiplier tubes (PMTs) and two detectors. The service (repair/replacement) history of these PMTs and detectors were available as unstructured, free-format text in the log files. Like the extraction phase of typical ETL systems, the goal was to convert this unstructured textual history into structured data, which gives accurate description of when each PMT was serviced/replaced, and which detector(s) was/were serviced alongwith it. We implemented a text analytic tool using the regular expression API of Java to accomplish this.

[1] "Efficient Elastic Burst Detection in Data Streams", Yunyue Zhu and Dennis Shasha, Proceedings of the 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2003
[2] "Bursty and Hierarchical Structure in Streams", Jon Kleinberg, Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2002