Publikationen

Semiconductor manufacturing is a highly innovative branch of industry, where a high degree of automation has already been achieved. For example, devices tested to be outside of their specifications in electrical wafer test are automatically scrapped. In this paper, we go one step further and analyze test data of devices still within the limits of the specification, by exploiting the information contained in the analog wafermaps. To that end, we propose two feature extraction approaches with the aim to detect patterns in the wafer test dataset. Such patterns might indicate the onset of critical deviations in the production process. The studied approaches are: 1) classical image processing and restoration techniques in combination with sophisticated feature engineering and 2) a data-driven deep generative model. The two approaches are evaluated on both a synthetic and a real-world dataset. The synthetic dataset has been modeled based on real-world patterns and characteristics. We found both approaches to provide similar overall evaluation metrics. Our in-depth analysis helps to choose one approach over the other depending on data availability as a major aspect, as well as on available computing power and required interpretability of the results.

Season length estimation is the task of identifying the number of observations in the dominant repeating pattern of seasonal time series data. As such, it is a common pre-processing task crucial for various downstream applications. Inferring season length from a real-world time series is often challenging due to phenomena such as slightly varying period lengths and noise. These issues may, in turn, lead practitioners to dedicate considerable effort to preprocessing of time series data since existing approaches either require dedicated parameter-tuning or their performance is heavily domain-dependent. Hence, to address these challenges, we propose SAZED: spectral and average autocorrelation zero distance density. SAZED is a versatile ensemble of multiple, specialized time series season length estimation approaches. The combination of various base methods selected with respect to domain-agnostic criteria and a novel seasonality isolation technique, allow a broad applicability to real-world time series of varied properties. Further, SAZED is theoretically grounded and parameter-free, with a computational complexity of O(𝑛log𝑛), which makes it applicable in practice. In our experiments, SAZED was statistically significantly better than every other method on at least one dataset. The datasets we used for the evaluation consist of time series data from various real-world domains, sterile synthetic test cases and synthetic data that were designed to be seasonal and yet have no finite statistical moments of any order.

Semiconductor manufacturing processes critically depend on hundreds of highly complex process steps, which may cause critical deviations in the end-product.Hence, a better understanding of wafer test data patterns, which represent stress tests conducted on devices in semiconductor material slices, may lead to an improved production process.However, the shapes and types of these wafer patterns, as well as their relation to single process steps, are unknown.In a first step to address these issues, we tailor and apply a variational auto-encoder (VAE) to wafer pattern images.We find the VAE's generator allows for explorative wafer pattern analysis, andits encoder provides an effective dimensionality reduction algorithm, which, in a clustering application, performs better than several baselines such as t-SNE and yields interpretable clusters of wafer patterns.

Modeling activity in online collaboration websites, such asStackExchange Question and Answering portals, is becom-ing increasingly important, as the success of these websitescritically depends on the content contributed by its users. Inthis paper, we represent user activity as time series and per-form an initial analysis of these time series to obtain a bet-ter understanding of the underlying mechanisms that governtheir creation. In particular, we are interested in identifyinglatent nonlinear behavior in online user activity as opposedto a simpler linear operating mode. To that end, we applya set of statistical tests for nonlinearity as a means to char-acterize activity time series derived from 16 different onlinecollaboration websites. We validate our approach by com-paring activity forecast performance from linear and nonlin-ear models, and study the underlying dynamical systems wederive with nonlinear time series analysis. Our results showthat nonlinear characterizations of activity time series helpto (i) improve our understanding of activity dynamics in on-line collaboration websites, and (ii) increase the accuracy offorecasting experiments.

This paper provides an overview of current literature on timeseries classification approaches, in particular of early timeseries classification.A very common and effective time series classification ap-proach is the 1-Nearest Neighbor classifier, with differentdistance measures such as the Euclidean or dynamic timewarping distances. This paper starts by reviewing thesebaseline methods.More recently, with the gain in popularity in the applica-tion of deep neural networks to the field of computer vision,research has focused on developing deep learning architec-tures for time series classification as well. The literature inthe field of deep learning for time series classification hasshown promising results.Early time series classification aims to classify a time se-ries with as few temporal observations as possible, whilekeeping the loss of classification accuracy at a minimum.Prominent early classification frameworks reviewed by thispaper include, but are not limited to, ECTS, RelClass andECDIRE. These works have shown that early time seriesclassification may be feasible and performant, but they alsoshow room for improvement