시계열 데이터에 대한 전이 학습 기반 이상 탐지 방법 연구

Abstract

Anomaly detection is crucial for ensuring the reliability and safety of mechanical systems across industries such as power generation, manufacturing, and transportation. In these mechanical systems, data is usually collected in time-series form using sensors such like vibration, current or sound for anomaly detection. Time-series anomaly detection methods often face limitations due to insufficient training data and poor generalization across complex operating conditions and varying loads. To address these challenges, this study proposes a transfer learning-based anomaly detection model, leveraging pre-trained knowledge to deliver robust performance and adaptability in data-scarce scenarios and diverse industrial environments. To this end, time-series signals are transformed into spectrograms through Short-Time Fourier Transform(STFT), followed by feature extraction through a Convolutional Autoencoder to obtain low-dimensional latent features. These features are used to detect anomalies using classification such as Random Forest and eXtreme Gradient Boosting. Building on this approach, this research validates the model's performance through migration tasks using the Case Western Reserve University(CWRU) Bearings dataset. Furthermore, to show cross-condition generalization, the proposed model was validated on the Hanoi University of Science and Technology(HUST), Sumair– Umar Bearing Fault(SUBF) dataset v2.0, and a dataset collected using microphone sensor in motor dynamo tests. Consequently, unlike other studies limited by specific operating conditions, the proposed model exhibits strong generalization performance across bench- mark datasets. Experimental results highlight the effectiveness of combining STFT, CAE, and tree-based classifiers in addressing data scarcity and enhancing generalization, making it highly suitable for real-world industrial applications. Future work will focus on noise-robust techniques and broader fault types to further improve performance.