A noise-robust bearing fault diagnosis approach for drying equipment using improved image transformation on convolutional neural networks

Real-time monitoring and diagnosis of bearings are vital for ensuring the normal operation of drying equipment. Rolling bearings often run under noise conditions, making their fault characteristics difficult to be recognized from vibration signals. A noise-robust bearing fault diagnosis approach was put forward and validated on convolutional neural networks. 1D vibration signals were first transformed into 2D color images by developing an improved image transformation method that combined data enhancement, fast Fourier transform, and threshold processing. The 2D color images were input into the modified AlexNet with reduced channels while adopting hyperparameter auto-tuning and image enhancement strategies for training to enhance the diagnosis efficiency and accuracy. The validations were conducted on the three datasets from Case Western Reserve University, Jiangnan University, and Southeast University at varying signal-to-noise ratios. The results indicated that the averaged diagnosis accuracies were 96.94, 98.29, and 99.82%, respectively, for the three datasets. The accuracies at −10 dB of the strong noise remained 91.59, 96.74, and 99.71%, and the accuracies under the noise-free condition all reached over 99.5% for the three datasets. Furthermore, the improved image transformation method was successfully extended to other convolutional neural networks, demonstrating its adaptivity and generality in enhancing the diagnosis accuracies. The proposed approach demonstrated strong potential for application in the monitoring of industrial drying equipment operating under noisy conditions.