Label-Free Power Failure Detection in CNC Machine Tools via Cross-Domain Energy Residual Monitoring
Abstract Predictive maintenance of CNC machine tools is limited by scarce fault-labelled data and the need to commission a model separately for each machine. This paper proposes CEAD (Cross-domain Energy Anomaly Detector), a label-free framework that requires neither fault labels nor target-domain training data. CEAD trains a Fuzzy Inference System and Extreme Gradient Boosting (XGBoost) energy model exclusively on 1,000 normal-operation records from a manufacturer-provided simulation dataset and freezes it. It then deploys the frozen model on a structurally different target (AI4I 2020, a synthetic milling-machine benchmark of 10,000 samples) through a deterministic feature mapping with no learnable parameters. The absolute difference between the predicted mechanical power and the normalised actual mechanical power defines the Energy Residual Index (ERI), a scalar health indicator in source-domain-referenced watts. A percentile control limit on a short calibration window of verified normal operation sets the alert threshold, so that no stage of the framework uses fault labels. Evaluated on 9,000 held-out AI4I 2020 samples (95 Power Failure events), CEAD achieves an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) = 0.9914 and, at the 95th-percentile threshold, recall = 93.7% at a false positive rate of 3.8%. It outperforms Isolation Forest, One-Class SVM, and Local Outlier Factor by at least 9.6 percentage points in AUC-ROC and misses 6 Power Failure events against 45–59 for the baselines. These results remain consistent for calibration windows as small as 500 samples. Power Failure elevates the mean ERI from 133 W to 363 W (2.72×; Mann–Whitney p < 0.001), while mechanical fault modes yield at most a moderate response (AUC-ROC ≤ 0.693). CEAD thus enables predictive maintenance deployment from manufacturer simulation data and a short verified-normal recording, without fault-labelling campaigns or per-machine model training.
Authors
- Tran Ngoc Hoang (ORCID: https://orcid.org/0000-0001-8683-0248)
Institutions
- University of Da Nang (VN)
- Trường Đại học Sư phạm Kỹ thuật (VN)
Publication Details
- Journal
- Engineering Research Express
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1088/2631-8695/aec36c
- Primary Topic
- Machine Fault Diagnosis Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Đại học Đà Nẵng