This research aims to address the challenge of limited labeled data in industrial fault detection. By leveraging pseudo-label learning and frequency-domain knowledge, PF-MCNN seeks to improve diagnostic accuracy while reducing reliance on costly manual labeling, thus enabling more practical deployment in real-world predictive maintenance.
This study develops PF-MCNN, a semi-supervised bearing fault detection method. The approach integrates convolutional neural networks (CNNs) with signal transformations in multiple frequency domains and introduces a pseudo-label filtering mechanism based on frequency similarity. Three CNN−transformer models capture complementary features from low, medium and high frequencies, and filtered pseudo-labels are iteratively refined to train the final model.
Experimental evaluation on the CWRU data set shows that PF-MCNN achieves 92.41% accuracy, outperforming traditional pseudo-labeling CNN methods. The model converges faster and demonstrates stronger robustness by effectively combining multi-frequency convolution kernels with dynamic pseudo-label thresholds guided by signal transformations.
PF-MCNN uniquely connects CNN convolution operations with signal transformation principles, enhancing interpretability and diagnostic power. Unlike conventional methods, it integrates physical signal properties into pseudo-label filtering, improving learning efficiency under scarce data conditions. This contribution offers both theoretical novelty and practical value for industrial fault detection systems.
