Denoisis: multibearing fault diagnosis based on time–frequency denoising autoencoder and global-local domain adapter for zero-shot learning and domain adaptation

Abstract Measurement-based bearing fault diagnosis with deep learning is an important problem in many industrial scenarios, and the focus has been switching from diagnosing single bearing faults to extracting fault features in the sensor measurements of multibearings in a system. To achieve effective multibearing diagnosis, two challenges need to be addressed. (1) It is difficult or impossible to obtain data of all combinations of faults for the multibearings. (2) There is, usually, some working condition difference between the source multibearing data for training and the target multibearing data for diagnosis. Different from single bearing faults, the multibearing fault data are correlated, that is, one bearing’s fault data are affected by other bearings as noises, whereas one bearing’s data could share some similarity to other bearings with the same fault type. To address the challenges associated with the above analysis, we propose Denoisis 6 6 The name Denoisis means denoising diagnosis. , a multibearing fault diagnosis based on time–frequency denoising autoencoder and global-local domain adapter for zero-shot learning and domain adaptation. Denoisis can be trained with single bearing fault data for multibearing fault diagnosis and thus attain zero-shot learning because of its time–frequency denoising autoencoder to remove the noise in multibearing fault data and the Transformer group with cross-bearing attention to utilize the similarity information from the data. Denoisis can achieve domain adaptation because of its global-local domain adapter with domain discrimination losses, a consistency loss and a voting loss. In our tests on the BJTU-RAO dataset, experimental results demonstrate that Denoisis is effective and accurate in multibearing diagnosis.