TFE-SWResNet: an improved residual network based on time–frequency feature enhancement for bearing fault diagnosis

Abstract

Rolling bearings are critical transmission components in rotating machinery. Their operating conditions directly determine the operational safety and overall reliability of the entire mechanical system. Traditional convolutional neural networks (CNNs) are limited in modelling long-range dependencies and high-order fault features. In particular, when applied to bearing fault classification tasks based on one-dimensional vibration signals, it is difficult for CNNs to effectively characterize the time-varying and non-stationary features of the signals, and they exhibit insufficient capability to extract high-order fault features. To resolve the issues outlined above, this paper proposes an improved residual network (ResNet) fault diagnosis model based on time–frequency feature enhancement (TFE). Using two-dimensional time–frequency maps as input, the proposed model performs adaptive feature reconstruction by employing quadratic nonlinear convolution in the TFE module, while enhancing the representation of nonlinear fault features by combining Hilbert envelope analysis and multi-norm normalization. Specifically, the TFE front end is integrated with a newly designed shifted-window ResNet (SWResNet) backend to form TFE-SWResNet. By integrating depthwise separable convolution with shifted-window multi-head self-attention within the residual architecture, the proposed model coordinates fine-grained local feature extraction with cross-window contextual modelling while maintaining controlled computational cost and stable feature propagation. To evaluate the effectiveness of the proposed model, comparative experiments were conducted on the Case Western Reserve University (CWRU) dataset, with samples pooled from four load-speed conditions, and on the PU dataset, which contains fault categories of varying complexity under a fixed operating condition. The experimental results show that the proposed model achieves average accuracies of 98.68% and 98.79% on the CWRU and PU datasets, respectively. Compared with five conventional baseline models and three additional related models, TFE-SWResNet demonstrates higher diagnostic accuracy and better classification stability across the two evaluated benchmark datasets.