Abstracts Track 2026


Area 1 - Industry of the Future

Nr: 25
Title:

Unsupervised Deep Learning Methods for Robust Electric Motor Fault Detection

Authors:

Jernej Mlinaric and Gregor Dolanc

Abstract: Reliable end-of-line (EoL) quality inspection is a key requirement for intelligent industrial production systems and Industry 4.0 manufacturing environments. Modern electric motor production heavily relies on data-driven inspection methods that can detect defects and faults under varying process and environmental conditions. However, conventional anomaly detection systems based on static thresholds are often less robust in case of production variability, sensor drift, or measurement noise. This work suggests an adaptive thresholding approach for unsupervised deep learning-based EoL quality inspection of electric motors. The proposed method uses a hybrid CNN–BiGRU autoencoder trained exclusively on healthy production data represented by Mel-frequency spectrograms (MFSs) derived from vibration and acoustic signals. Instead of using a fixed statistical decision boundary, the anomaly threshold is dynamically derived from the distribution of reconstruction errors using percentile-based adaptation. The method was evaluated on an industrial dataset of 2,402 electric motors collected from a real manufacturing environment. Experimental results show that the adaptive thresholding approach improves fault detection robustness and eliminates false negatives while maintaining low false-positive rates. Compared with fixed-threshold approaches, the proposed method achieved improved classification performance with a ROC-AUC of 0.995 and an F1-score of 0.9904. This approach is practical, scalable and cost-effective solution for intelligent industrial inspection systems. It enables robust and autonomous quality control without requiring labelled fault data or manual recalibration, supporting the development of adaptive Industry 4.0 production environments and data-driven manufacturing systems.