Proportional periodic sampling for cross-load bearing fault diagnosis

Bearing vibration data under various loads shows different distributions, which leads to the poor performance of existing deep learning methods in performing cross-load fault diagnosis tasks. As a result, cross-load deep transfer learning methods that can reduce the impact of distribution differences have become the current research priority. However, existing transfer learning methods fall short in fully leveraging the bearing data characteristics, resulting in unreasonable sampling strategy, complex or inefficient network structure. These limitations hamper the performance improvement of cross-load fault diagnosis methods. To address the above drawbacks, a transductive convolution transfer learning (TCTL) method based on proportional periodic sampling is proposed. First, according to the periodic characteristic of bearing vibration data, the proportional periodic sampling is conducted on all vibration data to construct high-quality sample sets, which can greatly reduce the number of model parameters and ensure the excellent diagnostic accuracy of the model. Second, considering the short length of vibration samples and minor distribution differences between various loads, the cross-load convolutional neural network (CL-CNN) model is proposed to extract fault features from the constructed samples efficiently. Third, utilizing the characteristic that the vibration data under each load satisfies the clustering assumption, the cross-load multi-objective loss is used to effectively supervise the CL-CNN model to extract more domain-invariant features. Finally, compared to ablation and latest methods, TCTL boasts higher average value and lower standard deviation of accuracy on both datasets, proving the effectiveness of TCTL.