A Bearing Fault Diagnosis Method Based on Vibration Signal Extension and Time–Frequency Information Fusion Network Under Small Sample Conditions

Due to the limited fault samples, the accuracy of the bearing fault diagnosis model is challenged. Therefore, this article proposes a bearing fault diagnosis method based on vibration signal extension (VSE) and time-frequency information fusion network (TFIF-Net). In the preprocessing stage, a VSE method is proposed. VSE predicts future trends based on changes in the existing signal, thereby extending the length of the signal and expanding the fault sample. Additionally, to prevent the loss of crucial features during extension, an extended optimization algorithm is proposed. By minimizing the square of the frequency error of each signal component before and after extension, this algorithm ensures that the spectrum distribution of the extended signal is similar to that of the original signal, thereby providing more reliable samples. In the diagnosis stage, a TFIF-Net is proposed. TFIF-Net uses the input data from both the time and frequency domains provided by VSE, fully extracting valuable fault features from multiple sources and accurately diagnosing bearing faults. Additionally, given the varied contributions of different domain features to the diagnosis, TFIF-Net enhances features that have a more substantial impact on the diagnosis by assigning distinct weights to time- and frequency-domain features. The superiority of the proposed method is verified on both a public dataset and an autonomously collected dataset. Specifically, when there are five original training samples for each type of working condition, the diagnostic accuracy is not less than 99.3%, which is significantly higher than that of the comparison methods.