A Physics-Informed Temporal Convolutional Network for Joint Fault Diagnosis and Calibrated Severity Estimation in Induction Motors Through Motor Current Signature Analysis

Mohammed, A ORCID: 0009-0003-8426-6964, Abdullah, BM ORCID: 0000-0002-1281-148X, Zhang, Q ORCID: 0000-0002-0651-469X, Aldhaibani, O ORCID: 0000-0003-0235-2862, Cullen, J ORCID: 0000-0003-0401-4613, Raschella, A ORCID: 0000-0002-1626-8947 and Salih, A ORCID: 0009-0007-1968-240X (2026) A Physics-Informed Temporal Convolutional Network for Joint Fault Diagnosis and Calibrated Severity Estimation in Induction Motors Through Motor Current Signature Analysis. Machines, 14 (10). pp. 1-31. ISSN 2075-1702

Preview

Text A Physics Informed Temporal Convolutional Network for Joint Fault Diagnosis.pdf - Published Version Available under License Creative Commons Attribution. Download (1MB) | Preview

Abstract

Motor current signature analysis (MCSA) enables non-invasive condition monitoring of three-phase induction motors, but most published studies use short laboratory records with artificially seeded faults and report only a discrete fault label without any measure of confidence. This paper proposes a physics-informed temporal convolutional network (PI-TCN) that is trained and tested on 12 months of stator-current records measured on a 90 kW industrial induction motor in service. Fault labels were assigned retrospectively by aligning the recordings with documented maintenance interventions. From each 10 s current snapshot, 60 features are computed that describe time-domain statistics, inter-phase symmetry, harmonic distortion, the energy at theoretical fault frequencies, and the shape of the spectrum. Using a 24 h history of these features, the network classifies 10 motor conditions, estimates a continuous fault severity between 0 and 1, and provides prediction intervals for the severity. To avoid optimistic results, the data are divided by fault episode, with a temporal gap between training and test data. On the independent test set, the PI-TCN reaches an accuracy of 0.991±0.002 and a macro-averaged F1 score of 0.898±0.009 over five training runs, and estimates the severity of faulty windows with a mean absolute error of 0.031±0.004. A five-member ensemble of the network gives prediction intervals whose empirical coverage matches the nominal level from 50% to 95% (expected calibration error 0.015), whereas split-conformal intervals under-cover at the low levels because calibration and test windows are not exchangeable under the temporal split. A label-perturbation study shows that a one-day error in the fault onset lowers the macro-averaged F1 by up to 0.11 or more than doubles the severity error, depending on its direction. Because the method uses only the existing plant data logger, it can be deployed without additional sensors.

Item Type:	Article
Uncontrolled Keywords:	40 Engineering; 4008 Electrical Engineering; 4009 Electronics, Sensors and Digital Hardware; 0903 Biomedical Engineering; 4003 Biomedical engineering
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science T Technology > TA Engineering (General). Civil engineering (General) T Technology > TJ Mechanical engineering and machinery
Divisions:	Computer Science and Mathematics Engineering and Built Environment
Publisher:	MDPI AG
Date of acceptance:	3 October 2026
Date of first compliant Open Access:	29 October 2026
Date Deposited:	29 Oct 2026 13:40
Last Modified:	29 Oct 2026 13:40
DOI or ID number:	10.3390/machines14101171
URI:	https://researchonline.ljmu.ac.uk/id/eprint/29764

View Item