온도 데이터 분석을 통한 냉동 창고의 이상 징후 탐지

초록

This study addresses the problem of anomaly detection in cold storage systems, specifically focusing on abnormalities during the defrosting cycle. Since continuous manual monitoring of storage conditions incurs significant costs, an automated monitoring system based on temperature data is proposed. Unlike traditional methods that rely on supervised learning with both normal and abnormal data, this study employs an unsupervised approach using only normal state data. Specifically, an LSTM-Autoencoder model is utilized to learn the temporal patterns of normal operations by compressing and reconstructing the input data. The model exhibits low reconstruction errors for data similar to the training set (normal) and high errors for anomalous data. Leveraging this characteristic, anomalies are detected based on the reconstruction error calculated via Mean Squared Error (MSE). To validate the model, four potential failure modes were defined, and the anomaly scores were evaluated. Considering the skewed distribution of the reconstruction errors from the normal data, the Median Absolute Deviation (MAD) was adopted to set the anomaly threshold, rather than relying on the mean and standard deviation. Although obtaining real-world abnormal data remains a challenge, this study demonstrates the feasibility of detecting system faults using only normal operational data. Future research will focus on validating and refining the model as more abnormal data becomes available.

키워드