Automatic fault diagnosis of rolling bearings under multiple working conditions based on unsupervised stack denoising autoencoder

In practical engineering, data often lack labels, resulting in difficulty in fault diagnosis. Because stack-denoising autoencoders possess robust feature extraction capabilities and resistance to interference, an automatic and unsupervised bearing fault diagnosis method based on the stack-denoising autoencoder without an output layer was proposed in this study. As the stacked denoising autoencoder is an unsupervised algorithm, this approach can reduce reliance on manually labeled data labels. Therefore, this study proposed a new method for automatic fault diagnosis. First, the bearing fault features of the rolling bearing were extracted using the stack denoising autoencoder without an output layer. Meanwhile, the dimensions of the features were directly reduced to two or three dimensions by several hidden layers, thereby reducing manual experience. Second, the labels extracted from the clustering model were selected as inputs for different classifier models to automatically identify different types of faults. Two open-source rolling bearing datasets under various conditions were used to validate the classification performance of the proposed method. Finally, its effectiveness was verified using the experimental results. Various indicators were used to evaluate the performance of the proposed method, and the results showed an automatic bearing fault diagnosis accuracy of up to 90% when using different models and working conditions. Among the two datasets, the classification model achieved the highest accuracies of 0.99667 and 0.97143 and the lowest accuracies of 0.98000 and 0.90476, respectively.