A Hybrid Time-Frequency and Lightweight Deep Learning Model for Intelligent Fault Diagnosis of Rolling Bearings

A Hybrid Time-Frequency and Lightweight Deep Learning Model for Intelligent Fault Diagnosis of Rolling Bearings driven CNNs for bearing fault diagnosis is validated, laying a foundation for optimizing and improving fault diagnosis methods for various scenarios. However, the performance of CNNs heavily relies on the quality and diversity of training data, and their effectiveness under strong noise interference or imbalanced sample conditions remains limited. (2) A method named CWT-Mobile Net is proposed to address the issue of low recognition accuracy for certain fault types in practical scenarios due to noise interference and imbalanced fault samples. This method combines continuous wavelet transform (CWT) and lightweight neural network MobileNet-v2. Initially, CWT is applied using the Cmor wavelet basis function tailored for vibration signals to transform one-dimensional vibration signals into two-dimensional time-frequency maps. These maps serve as inputs to the lightweight MobileNet-v2 network, reducing feature information loss and suppressing noise interference. Experimental results validate the effectiveness of CWT-Mobile Net in achieving good recognition accuracy for various fault types while achieving network light weighting through optimizations such as depth-wise separable convolution, inverted residual blocks, and linear bottleneck structures. It should be noted that the current method utilizes a fixed Cmor wavelet basis, which may limit adaptability to signals with non-impulsive characteristics. Furthermore, the experiments were conducted under controlled noise levels and specific operating conditions (CWRU dataset), and its generalization capability for industrial scenarios with extreme noise or variable-speed conditions requires further verification.