A multi-level adaptation scheme for hierarchical bearing fault diagnosis under variable working conditions

Bearing fault diagnosis is important during the operation of mechanical equipment. Traditional deep-learning -based methods afford excellent diagnostic results if the training and test samples are of similar distribution. Thus, the datasets used for training and testing are collected under the same working conditions. However, when the working conditions change, a fault diagnosis model trained using such a training set cannot be directly applied to the test set. In addition, existing classification methods ignore the hierarchical structure of bearing fault categories, thus treating all categories equally. To address these issues, we present a new form of branch multi-level adaptation based on a convolutional neural network (BMACNN model). A branch structure is added to a one-dimensional CNN to permit hierarchical diagnosis of bearing faults via multiple output layers. The multiple kernel variant of the maximum mean discrepancy is used to regularize the loss function, thus reducing distributional distances among the domains of multiple levels. We tested the BMACNN model using six distinct transfer fault diagnostic scenarios of the Paderborn dataset. The BMACNN robustly adapted to variable working conditions and was superior to other methods.

Automated summary

Bearing fault diagnosis is crucial for the operation of mechanical equipment. Traditional deep-learning methods perform well when training and testing samples have similar distributions, but they fail to adapt to changing conditions. To address this, a new branch multi-level adaptation model based on convolutional neural networks is proposed, enabling hierarchical diagnosis of bearing faults and reducing distributional distances among multiple levels. Experimental results show its robust adaptability and superior performance compared to other methods.