A Novel Approach for Induction Motor Bearing Fault Detection Using Motor Current Signals and Quantum Self-Attention Neural Network

Induction motors are widely employed across industrial applications because of their robustness and simplicity. Ensuring their reliable and continuous operation necessitates the early detection of electrical and mechanical faults. However, the reliance on external sensors for collecting vibration signals presents a significant drawback, as their installation can be challenging and costly, particularly in hard-to-reach areas. This paper proposes a novel technique for induction motor bearing fault detection utilizing motor current signals and a quantum self-attention neural network (IMBFD-MCS-QSANN). The input data are collected and then processed using a Risk-Sensitive Extended Kalman Filter (RSEKF) for pre-processing. The RSEKF is used for data filtering, data normalization and data decimation. Then, the pre-processed data are given to the Synchrosqueezing Fractional Wavelet Transform (SFWT) for feature extraction. Standard deviation, variance, mean, median and other statistical variables are extracted using SFWT. The extracted features are input to the Quantum Self-Attention Neural Network (QSANN) for fault diagnosis, which classifies the bearing conditions into healthy, outer race fault, or inner race fault. The network’s weight parameters are optimized using the Black-Winged Kite Algorithm (BWKA) to develop classification accuracy. The performance of the proposed IMBFD-MCS-QSANN method was evaluated utilizing several metrics, including Accuracy, Precision, [Formula: see text]-score, Recall, Specificity, Receiver Operating Characteristic (ROC) and Loss. The method demonstrates outstanding results across all fault categories, achieving 99.73% for healthy bearings, 98.1% for outer race faults and 95.3% for inner race faults, with an overall accuracy of 99.73%. These outcomes highlight the superior effectiveness of the technique compared to existing approaches, such as the IoT-Integrated Multi-Parallel Graph Convolutional Network with frequency attention (MBFD-MGCN) and the Graph Neural Network-based method utilizing multi-relationships of intrinsic mode functions for multiple mechanical faults (IMF-MMF-GNN).