차량부품 내구평가 데이터를 활용한 딥러닝 기반 다중 클래스 이상탐지 모델 구현

다국어 초록 (Multilingual Abstract)

The automotive industry's increasing demand for reliability and durability has led to rigorous testing processes, including endurance tests where components are operated until failure. Traditional anomaly detection methods based on signal thresholds face challenges: high thresholds can cause secondary failures or distort cycle times, while low thresholds trigger premature interventions. To overcome these limitations, this study employs a dual LSTM-AE (Long Short-Term Memory Autoencoder) model, which excels at capturing time-series data patterns and detecting anomalies through reconstruction errors. By utilizing two LSTM-AE models, this approach allows for a more precise classification of operational states into normal, caution, and critical zones. Using the Power Take-off Unit (PTU) as a case study, the dual LSTM-AE model provides actionable insights to prevent secondary failures. The paper discusses the test environment, data preprocessing, model design, and anomaly detection results.