A new meta-transfer learning method with freezing operation for few-shot bearing fault diagnosis

Abstract

Deep learning (DL) for bearing fault diagnosis often requires a large quantity of comprehensive data to support in the field of rotating machinery fault diagnosis. However, large-quantity dataset for model training is difficult to obtain in the actual working environment. Therefore, bearing fault diagnosis problems under practical working conditions are often few-shot problems. Meta-learning can be adopted to solve few-shot problems. Traditional meta-learning method will lead to model overfitting, and shallow neural networks are usually used to avoid overfitting. As a result, features extracted by the shallow neural network are insufficiently rich to exploit the optimal performance of the model. A few-shot fault diagnosis method based on meta-learning named meta-transfer learning method with freezing operation (MTLFO) is proposed in this study to solve these problems. MTLFO can learn new knowledge rapidly through a small number of samples. Hyperparameter self-regulation ability of meta-learning is adopted by MTLFO, and freezing operation (FO) is used to deal with the neuronal of meta-learning to ensure that neurons from different tasks are transferred in the manner of scaling and shifting. MTLFO avoids the overfitting problem in traditional meta-learning methods and presents more advantages in solving few-shot problems in fault diagnosis compared with other groups of methods.

Transfer multiscale adaptive convolutional neural network for few-shot and cross-domain bearing fault diagnosis

“…To validate the superiority of the proposed TMACNN, we further conducted few-shot and cross-domain fault diagnosis comparison experiments on the CWRU dataset and the Paderborn dataset, using four baseline models, namely TCNN [33], T-AWMSCNN [3], T-WDCNN [34] and MTLFO [12]. TCNN is an advanced transfer CNN framework that enhances diagnostic accuracy and computational efficiency under As shown in figure 12, the proposed TMACNN can effectively leverage knowledge from the source domain to achieve target domain transfer fault diagnosis.…”

“…However, few-shot learning strategies can get into trouble when the training dataset are extremely scarce. Wang et al [12] proposed a meta-transfer learning with freezing operation (MTLFO) method for fewshot fault diagnosis, and it avoids the overfitting problem in traditional meta-learning methods.…”

Transfer learning rolling bearing fault diagnosis model based on deep feature decomposition and class-level alignment

“…Consequently, there is a proposition to apply transfer learning and metalearning to the field of bearing fault diagnosis. Wang et al [25] introduced a few-shot fault diagnosis approach based on meta-learning called 'meta-transfer learning with frozen operations' (MTLFOs). MTLFO leverages the self-adaptation capability of meta-learned hyperparameters and employs frozen operations on neurons to ensure the transfer of neurons across different tasks through scaling and shifting.…”

A meta-learning method for few-shot bearing fault diagnosis under variable working conditions

“…Li et al [1] focused on treating fault modes under different conditions as fine-grained faults, employing an attention mechanism in a three-stage learning process. Wang et al [25] proposed meta-transfer learning with a freezing operation, which involves leveraging hyperparameter selfregulation and freezing operations to address neuronal properties in meta-learning and prevent overfitting. Prototypical networks, a metric-based meta-learning strategy, achieve classification by calculating the Euclidean distances between sample features and class prototypes, assessing the similarity between support and query samples [26].…”