TY  - EJOU
AU  - Alamri, Faten S. 
AU  - Saleem, Adil Ali 
AU  - Khan, Muhammad I. 
AU  - Siddiqui, Hafeez Ur Rehman 
AU  - Rehman, Amjad 

TI  - Noninvasive Radar Sensing Augmented with Machine Learning for Reliable Detection of Motor Imbalance
T2  - Computer Modeling in Engineering \& Sciences

PY  - 2026
VL  - 146
IS  - 1
SN  - 1526-1506

AB  - Motor imbalance is a critical failure mode in rotating machinery, potentially causing severe equipment damage if undetected. Traditional vibration-based diagnostic methods rely on direct sensor contact, leading to installation challenges and measurement artifacts that can compromise accuracy. This study presents a novel radar-based framework for non-contact motor imbalance detection using 24 GHz continuous-wave radar. A dataset of 1802 experimental trials was sourced, covering four imbalance levels (0, 10, 20, 30 g) across varying motor speeds (500–1500 rpm) and load torques (0–3 Nm). Dual-channel in-phase and quadrature radar signals were captured at 10,000 samples per second for 30-s intervals, preserving both amplitude and phase information for analysis. A multi-domain feature extraction methodology captured imbalance signatures in time, frequency, and complex signal domains. From 65 initial features, statistical analysis using Kruskal–Wallis tests identified significant descriptors, and recursive feature elimination with Random Forest reduced the feature set to 20 dimensions, achieving 69% dimensionality reduction without loss of performance. Six machine learning algorithms, Random Forest, Extra Trees Classifier, Extreme Gradient Boosting, Categorical Boosting, Support Vector Machine with radial basis function kernel, and k-Nearest Neighbors were evaluated with grid-search hyperparameter optimization and five-fold cross-validation. The Extra Trees Classifier achieved the best performance with 98.52% test accuracy, 98% cross-validation accuracy, and minimal variance, maintaining per-class precision and recall above 97%. Its superior performance is attributed to its randomized split selection and full bootstrapping strategy, which reduce variance and overfitting while effectively capturing the nonlinear feature interactions and non-normal distributions present in the dataset. The model’s average inference time of 70 ms enables near real-time deployment. Comparative analysis demonstrates that the radar-based framework matches or exceeds traditional contact-based methods while eliminating their inherent limitations, providing a robust, scalable, and noninvasive solution for industrial motor condition monitoring, particularly in hazardous or space-constrained environments.
KW  - Condition monitoring; imbalance detection; industrial applications; machine learning; motor fault diagnosis; non-contact sensing; radar sensing; vibration monitoring

DO  - 10.32604/cmes.2025.074679