SpecCSAM: a prior-knowledge embedding deep learning fault diagnosis method for unknown working conditions

Abstract

Although deep learning has exhibited promising performance in the field of fault diagnosis, most current methods suffer from performance degradation under variable working conditions. Transfer learning is commonly used to address this problem, which assumes that a plenty of unlabeled data or a few labeled data can be obtained in target domains. However, in industrial scenarios, there are a number of unknown working conditions with the lack of data, leading to the failure of transfer learning. Therefore, this paper presents a prior-knowledge embedding deep learning fault diagnosis method, SpecCSAM to overcome this issue. Only the samples under the source working condition are utilized for model training, without using any data under unknown working conditions. Firstly, based on the bearing fault mechanism and spectrum analysis, a novel data preprocessing method is introduced to construct a spectrum feature matrix (SFM). This method highlights the features that are insensitive to the domain shift, which can effectively enhance the fault diagnosis performance under unknown working conditions. Secondly, in the feature extraction module, a fusion-encoder based on the multi-branch self-attention mechanism (SAM) is employed to capture the high-dimensional fault features in SFM. Based on this innovative SAM, it is capable of mining multi-scale correlation features. To demonstrate the superiority of the proposed method, experiments were carried out on the Case Western Reserve University dataset and the self-built dataset. The average accuracies under unknown working conditions reached 99.50 and 99.28%, respectively. Combined with other experimental results, the proposed method demonstrated excellent performance in multiple tasks under unknown working conditions.