Lightweight mobile inverted bottleneck convolution network with time–frequency Gramian angular field for bearing fault diagnosis

Abstract

Real-time detection of motor faults is crucial for timely maintenance, minimizing downtime, and preventing potential operational losses. Most existing studies have leveraged time–frequency (T–F) images in combination with deep learning techniques, which generally provide strong diagnostic performance and good generalization. However, these approaches often require computationally expensive preprocessing and incur high training overhead due to the extraction of complex features from T–F vibration images, thereby limiting their deployment in resource-constrained environments. To overcome these limitations, this work proposes a bearing fault diagnosis method using a lightweight MBConv network, combined with time–frequency Gramian Angular Field (T–F GAF) representations. This approach enables accurate classification of bearing faults while reducing computational complexity. The T–F GAF images generated from the time- and frequency-domain vibration signals of the Case Western Reserve (CWRU), Paderborn bearing and National Renewable Energy Laboratory (NREL)datasets were used as inputs to the proposed model for classification under different load conditions. Furthermore, ablation studies were conducted to evaluate the model’s overall effectiveness, the contribution of individual components, and the impact of optimization strategies. Performance was assessed using metrics such as F1-score, Recall, Precision, and the Confusion Matrix. The proposed model achieved a classification accuracy of 99.6% on CWRU, 98.15% on Paderborn bearing and 98.58% on NREL dataset under diverse load conditions with requiring lesser computational memory with an inference time 2 ms. In comparison, MobileNet-V2, MobileViT and LightVegNet achieved accuracies of 97% and 96% on different datasets, demonstrating the superior performance of the proposed approach.