Composite Score-Based Unsupervised Anomaly Detection Framework for Heat Treatment Process Temperature Data without Failure Labels

This study developed a composite score-based unsupervised temperature anomaly detection (CS-UTAD) framework for heat treatment process data without failure labels, maintenance records, or quality-defect labels. CS-UTAD reformulates the task as anomaly candidate screening and warning/alarm generation rather than supervised failure prediction. The framework combines forecasting residuals, autoencoder reconstruction, and statistical anomaly scores through rank fusion, supported by leakage-safe batch-wise splitting, train/validation-based threshold calibration, and a confirmed warning/alarm policy. On 345,390 one-second process records, CS-UTAD reduced the warning/alarm rate from 5.738% to 4.996% and the alarm rate from 2.971% to 2.531% compared with the guidebook autoencoder baseline; in the test interval, the corresponding rates were 4.986% and 2.613%. Its top-100 candidates included seven IQR weak-signal hits, whereas the baseline included none. The confirmed policy compressed 93,102 test samples into 51 alert episodes for expert review. The results support label-free anomaly monitoring rather than supervised failure prediction.