Research on the Role of MCP Regularization in Bearing Fault Diagnosis

In practical applications, the collection of bearing fault data often includes a significant number of invalid samples. When these data are used to train deep learning-based bearing fault diagnosis models, there is a risk of overfitting. Traditional methods, such as L1 and L2 regularization, are commonly employed to mitigate this risk. This paper proposes a novel regularization method based on the Minimax Concave Penalty (MCP) for deep learning models. Experimental validation was conducted using bearing fault data to train these models. Among the three groups of experiments, the highest average fault diagnosis accuracy after MCP regularization was 0.972. The results indicate that the MCP method demonstrates asymptotic normality in updating network weights and exhibits dynamic adaptability in penalizing the loss function. Therefore, the method proposed in this paper offers superior overfitting prevention in deep learning applications compared to L1 and L2 regularization.