A novel data-driven method based on sample reliability assessment and improved CNN for machinery fault diagnosis with non-ideal data

Recently, numerous new data-driven methods have been proposed. But most of them focused on the innovation of models and algorithms, and rarely discussed and optimized from the perspective of data and samples. However, the reliability of sample quality directly determines the effectiveness of machine learning models. In this paper, a novel data-driven method based on sample reliability assessment (SRA) and improved convolutional neural network (ICNN) for mechanical fault diagnosis was designed. First, multinomial logistic regression (MLR) was conducted to construct the assessment model and a statistical approach named influence function was used to compute the sample weights efficiently. Then, ICNN with the improved loss function was proposed based on the strategies of sample weights, class weights and early-stopping. Compared with traditional deep learning models, ICNN can better eliminate the negative impact of the problems during the model training including sample quality imbalance, class imbalance, and overfitting phenomenon. Therefore, the fault diagnosis performance can be improved. Finally, the trained ICNN can automatically extract the fault characteristics and achieve the fault diagnosis with the input of compressed time–frequency images. Experiments on a benchmarking dataset and a gear dataset from a practical experimental platform verified the superiority of the proposed fault diagnosis method.