DeepNAP: Deep neural anomaly pre-detection in a semiconductor fab

Chunggyeom Kim, Jinhyuk Lee, Raehyun Kim, Youngbin Park, Jaewoo Kang

    Research output: Contribution to journalArticlepeer-review

    28 Citations (Scopus)


    Anomaly detection in an industrial process is crucial for preventing unexpected economic loss. Among various signals, multivariate time series signals are one of the most difficult signals to analyze for detecting anomalies. Moreover, labels for anomalous signals are often unavailable in many fields. To tackle this problem, we present DeepNAP which is an anomaly pre-detection model based on recurrent neural networks. Without any annotated data, DeepNAP successfully learns to detect anomalies using partial reconstruction. Furthermore, detecting anomalies in advance is essential for preventing catastrophic events. While previous studies focused mainly on capturing anomalies after they have occurred, DeepNAP is able to pre-detect anomalies. We evaluate DeepNAP and other baseline models on a real multivariate dataset generated from a semiconductor manufacturing fab. Compared with other baseline models, DeepNAP achieves the best performance on both the detection and pre-detection of anomalies.

    Original languageEnglish
    Pages (from-to)1-11
    Number of pages11
    JournalInformation Sciences
    Publication statusPublished - 2018 Aug

    Bibliographical note

    Funding Information:
    This research was supported by Samsung Electronics, Co., Ltd. (No. GH170306 0003) and National Research Foundation of Korea (No. 2017R1A2A1A17069645).


    • Anomaly detection
    • Long short term memory
    • Multivariate
    • Time series data

    ASJC Scopus subject areas

    • Software
    • Control and Systems Engineering
    • Theoretical Computer Science
    • Computer Science Applications
    • Information Systems and Management
    • Artificial Intelligence


    Dive into the research topics of 'DeepNAP: Deep neural anomaly pre-detection in a semiconductor fab'. Together they form a unique fingerprint.

    Cite this