Deep Learning-Based Anomaly Detection in Electric Motor Production Using Mel-Spectrogram Representations and a Hybrid Neural Network

Abstract

End-of-Line (EoL) quality inspection of electric motors requires reliable detection of manufacturing faults before products leave the production line. However, conventional supervised deep learning approaches depend on a sufficient number of labeled instances of faulty motors, which are scarce in high-quality manufacturing environments. This limits their applicability to newly introduced products and previously unobserved fault types. This paper presents an unsupervised deep learning framework for anomaly detection in motor acoustic and vibration signals, designed for EoL quality inspection in manufacturing. The proposed method leverages Mel-frequency spectrograms (MFSs) as input features and employs a hybrid neural network combining a convolutional neural network (CNN) and bidirectional gated recurrent units (BiGRUs), effectively capturing both local spectral patterns and temporal dependencies. The method is evaluated on real industrial production data from over 2400 motors, of which approximately 4.3% were faulty, reflecting the highly imbalanced nature of high-quality manufacturing. The model was trained exclusively on healthy motor data and therefore does not require labeled faulty samples during training. Experimental results demonstrate strong discrimination between healthy and faulty motors, indicating that the proposed approach is suitable for automated EoL quality inspection in manufacturing environments where labeled faulty data are scarce.