An Unsupervised Variable Operating Condition Intelligent Diagnosis Method Based on Fault Feature Recognition in Time–Frequency Images of Rolling Bearings

Monitoring the health of rolling bearings is essential for stable industrial operation and cost reduction. However, traditional convolutional neural networks (CNNs) struggle to extract periodic fault features due to inadequate consideration of interfeature correlation. Furthermore, models reliant on single operating condition data face challenges in identifying fault modes across various conditions. To address these issues, this article introduces the Swin Transformer and conditional adversarial domain adaptation model for diagnosing rolling bearing faults under diverse operating conditions. The model employs continuous wavelet transform to convert vibration signals into 2-D time–frequency maps and utilizes windowing with sliding self-attention to enhance the extraction of periodic fault texture features and improve pattern recognition decision-making. Through innovative conditional adversarial domain adaptation and fault mode prediction assistance, the model effectively mitigates domain differences, enhancing the generalization of bearing health diagnosis across operational contexts. Testing on the Case Western Reserve University, Paderborn datasets, and a nonpublic dataset yielded average diagnostic accuracies of 99.92%, 95.4%, and 97.3%, respectively, outperforming CNNs and their variants. The model’s ability to visualize features confirms its capacity to capture periodic fault characteristics, indicating reliable diagnostic confidence and high practical value for industrial applications.