Unknown examples that are unseen during training often appear in real-world pattern recognition tasks, and an intelligent self-learning system should be able to distinguish between known examples and unknown examples. Accordingly, open-set recognition (OSR), which addresses the problem of classifying knowns and identifying unknowns, has recently been highlighted. However, conventional deep neural networks (DNNs) using a softmax layer are vulnerable to overgeneralization, producing high confidence scores for unknowns. In this article, we propose a simple OSR method that is based on the intuition that the OSR performance can be maximized by setting strict and sophisticated decision boundaries that reject unknowns while maintaining satisfactory classification performance for knowns. For this purpose, a novel network structure, in which multiple one-vs-rest networks (OVRNs) follow a convolutional neural network (CNN) feature extractor, is proposed. Here, an OVRN is a simple feedforward neural network that is designed to assign confidence scores that are lower than those in the softmax layer to unknown samples so that unknown samples can be more effectively separated from known classes. Furthermore, the collective decision score is modeled by combining the multiple decisions reached by the OVRNs to alleviate overgeneralization. Extensive experiments were conducted on various datasets, and the experimental results show that the proposed method performs significantly better than the state-of-the-art methods by effectively reducing overgeneralization. The code is available at https://github.com/JaeyeonJang/Openset-collective-decision.
9 November 2021 in IEEE Transactions on Industrial Informatics
In many real-world prognostics and health management tasks, where the available training samples are insufficient, deep neural networks are highly vulnerable to overfitting. To address this problem, in this article, we propose a novel health representation learning method based on a Siamese network. This method prevents overfitting by utilizing a constraint by which the differences between samples in the embedding space of the Siamese network should follow the differences in the remaining useful life (RUL) values via the introduction of a multitask learning scheme. In addition, since the learned embedding space reflects the dynamics of degradation, each training sample can be used as a reference to estimate the RUL of a test sample. By combining the estimates for all training samples, the proposed method enables robust RUL prediction. Experimental results show that the proposed learning and estimation method contributes to improving not only RUL prediction performance but also robustness to data insufficiency.
27 July 2020 in IEEE Transactions on Semiconductor Manufacturing
Wafer defect maps have different generation mechanisms according to the defect pattern, and automatic classification of wafer maps is therefore critical to reveal the root cause of the defects. In this paper, we examine the open set recognition problem, in which not only must wafer maps be classified using major defect patterns that are already known but also unknown defect patterns must also be detected. Our model is an ensemble model of a one-versus-one method that uses a convolutional neural network as the base classifier for wafer map classification. The proposed model calculates a weighted mean score for each defect pattern and determines the presence or absence of a pattern based on this score. The weight is calculated based on the proximity of data groups in the feature space and can be considered a support level at which a new wafer map belongs to a specific defect pattern. An untrained wafer map input into the model has a low support level and thus does not belong to any known defect pattern. An experiment was conducted using work-site failure bit count maps.
6 September 2024 in Expert Systems with Applications
The detection of wafer faults in early process steps through monitoring and analyzing multivariate process trace data contribute to wafer yield improvements. Standard classification algorithms have been generally used for fault detection and classification (FDC). However, this approach can cause information loss while extracting statistical features from the trace data and cannot consider class imbalance situations where much fewer faulty wafers are generated than normal wafers. In addition, the approach does not consider normal wafer-to-wafer (W2W) variations and sensor noise inherent in the trace data. These drawbacks significantly degrade FDC performance. This study proposes a method that builds an FDC model only with trace data of normal wafers in which W2W variations and sensor noise exist. The one-class FDC method detects the occurrence of abnormal trace patterns that cause wafer faults by removing W2W variations and sensor noise from raw traces by using denoising autoencoders, and this method finds the fault-introducing process parameters with the occurrence times. In experiments using the trace data of etch and chemical vapor deposition processes, the proposed method exhibited 1% and 6% higher performance than the best-performing method among comparison methods in terms of the geometric mean of the normal and fault detection accuracies.