In a semiconductor fabrication plant, various types of sensors are installed at various equipment of various processes to monitor the quality of the products and the progress of various processes. These sensors generate a huge volume of time-series waveform data, which are used for failure prognostics, failure diagnosis, and anomaly detection using supervised or unsupervised classifiers. For this purpose, automatic extraction of features from a huge volume of waveform data is needed; however, extraction of effective features automatically from a compound waveform having multiple spikes, complex states or transitions is difficult. If a compound waveform is properly segmented into multiple sub-waveforms, effective features from sub-waveforms can be extracted. In this paper, we propose a new waveform segmentation method based on two-step state and change-point detection using Kernel Density Estimation (KDE) and clustering techniques and apply IEEE Standard-based methods with our state determination technique to extract features from sub-waveforms. We apply our proposed method to waveform data of real sensors installed at one of our semiconductor fabrication plants and compare the performance with the conventional step-based segmentation technique and another pattern matching technique. The experimental results show that our proposed method enables extraction of effective features, which result in higher accuracy of anomaly detection compared to conventional techniques.

Sparse regression such as the Lasso has achieved great success in handling high-dimensional data. However, one of the biggest practical problems is that high-dimensional data often contain large amounts of missing values. Convex Conditioned Lasso (CoCoLasso) has been proposed for dealing with high-dimensional data with missing values, but it performs poorly when there are many missing values, so that the high missing rate problem has not been resolved. In this paper, we propose a novel Lasso-type regression method for high-dimensional data with high missing rates. We effectively incorporate mean imputed covariance, overcoming its inherent estimation bias. The result is an optimally weighted modification of CoCoLasso according to missing ratios. We theoretically and experimentally show that our proposed method is highly effective even when there are many missing values.

Shapelets are time-series segments effective for classifying time-series datasets. In recent years, the discovery of shapelets by classifier learning has been studied. Methods for shapelet discovery have attracted great interest because they provide not only interpretable results but also superior classifi-cation performance. However, they do not consider imbalanced classifications between majority and minority classes, which may occur in actual applications (e.g., anomaly detection). Our aim is to learn shapelets and classifiers using only training data for the majority class without the minority class. We propose a method called one-class learning time-series shapelets (OCLTS). OCLTS efficiently and simultaneously optimizes both the shapelets and a non-linear classifier based on a one-class support vector machine by a stochastic sub-gradient descent algorithm. Experimental results show the method's effectiveness for interpretability and imbalanced binary classification.