A Comparative Study of CNN, Multi-Head Self-Attention, and Hybrid CNN-MHSA Models for Supply-Voltage Prediction Using Motor Vibration Signals

Fatemeh Ghanadian, Amin Zollanvari, Askat Kural, Mehdi Bagheri

IEEE Access, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1109/access.2026.3708777
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Vibration-based condition monitoring is widely used to assess the health of electric motors. This study evaluates deep learning models for estimating motor supply-voltage levels using multichannel vibration measurements. The main objective is to investigate how different deep learning architectures capture vibration patterns associated with varying supply-voltage operating conditions and to evaluate their ability to infer operating-condition information directly from vibration measurements. In this context, we analyze a hybrid architecture that integrates Convolutional Neural Networks (CNN) with Multi-Head Self-Attention (MHSA), allowing the model to simultaneously learn local temporal patterns and long-range dependencies within the signal. Rather than proposing a single architecture optimized for a specific dataset, This study systematically compares three architectural frameworks: CNN, MHSA, and CNN-MHSA for multivariate time-series regression, where supply-voltage level is estimated from multichannel vibration measurements. The experiments are designed to facilitate a controlled comparison among these architectures. The proposed frameworks are evaluated using Relative Absolute Error (RAE), and a one-sided Wilcoxon signed-rank test is applied to examine statistical significance. The CNN-MHSA model achieved the best average test performance across five independent runs, with a mean RAE of $34.705 \pm 1.448$ , compared with $40.262 \pm 3.198$ for CNN and $73.069 \pm 7.472 $ for MHSA. The observed improvements were further supported by a one-sided Wilcoxon signed-rank test under the limited five-run experimental setting. This performance shows that local feature extraction with global context modeling can improve the representation of vibration responses, particularly under different motor supply-voltage and operating conditions.