A High-Precision voiceprint recognition for fault diagnosis of motor bearings through lightweight dynamic convolutional neural network

• Acoustic signals enable simple, non-contact, and low-cost fault diagnosis. • Experimental data of motor bearing faults are collected under realistic conditions. • The dynamic convolutional layer re-extracts key features for fault classification. • ADC-MobileNetV2 achieves 99.42% accuracy across five bearing fault types. Induction motors are essential components in industrial applications, and bearing faults are a common form of mechanical damage that can significantly reduce their efficiency. Non-contact acoustic fault diagnosis is a cost-effective and non-invasive method, but it is highly affected by environmental noise, which can alter the signal patterns related to each fault and obscure valuable information, and similarity in emitted sound patterns can reduce detection accuracy. Traditional methods have only attempted to reduce the effect of noise without robustifying the model against the noise distribution in acoustic features. To overcome this limitation, a novel design as an Attention-guided Dynamic Convolution (ADC) layer is introduced into the feature extraction stage of convolutional neural network (CNN) models, which enables the model to focus on specific patterns and regions that are related to faults. To evaluate the proposed approach, acoustic signals with background noise from five types of motor bearing faults were collected under experimental conditions and converted into Mel voiceprint inputs, demonstrating the superiority of the method with high computational efficiency, achieving 99.42% accuracy for the ADC-MobileNetV2 model with a total cost of 0.3 GFLOPs and an average inference time of 2.086 ms per sample.