No abstract

Rolling bearings play a pivotal role in rotating machinery. The remaining useful life prediction and fault diagnosis of bearings are crucial to condition-based maintenance. However, traditional data-driven methods usually require manual extraction of features, which needs rich signal processing theory as support and is difficult to control the efficiency. In this study, a bearing remaining life prediction and fault diagnosis method based on short-time Fourier transform (STFT) and convolutional neural network (CNN) has been proposed. First, the STFT was adopted to construct time-frequency maps of the unprocessed original vibration signals that can ensure the true and effective recovery of the fault characteristics in vibration signals. Then, the training time-frequency maps were used as an input of the CNN to train the network model. Finally, the time-frequency maps of testing signals were inputted into the network model to complete the life prediction or fault identification of rolling bearings. The rolling bearing life-cycle datasets from the Intelligent Management System were applied to verify the proposed life prediction method, showing that its accuracy reaches 99.45%, and the prediction effect is good. Multiple sets of validation experiments were conducted to verify the proposed fault diagnosis method with the open datasets from Case Western Reserve University. Results show that the proposed method can effectively identify the fault classification and the accuracy can reach 95.83%. The comparison with the fault diagnosis classification effects of backpropagation (BP) neural network, particle swarm optimization-BP, and genetic algorithm-BP further proves its superiority. The proposed method in this paper is proved to have strong ability of adaptive feature extraction, life prediction, and fault identification.

Conjugate gradient methods are a class of important methods for solving linear equations and nonlinear optimization problems. In this work, we propose a new stochastic conjugate gradient algorithm with variance reduction (CGVR 1 ), and we prove its linear convergence with the Fletcher and Reeves method for strongly convex and smooth functions. We experimentally demonstrate that the CGVR algorithm converges faster than its counterparts for four learning models, which may be convex, nonconvex or nonsmooth. Additionally, its area under the curve (AUC) performance on six large-scale datasets is comparable to that of the LIBLINEAR solver for the L2-regularized L2-loss but with a significant improvement in computational efficiency.

No abstract

No abstract