A New Robust Projection Distributed Broad Learning Under Redundant Samples and Noisy Environment

Broad learning system (BLS) is a breadth-based learning algorithm based on single-layer feedforward network (SLFN), which has the advantages of incremental learning with its fast-training speed and strong generalization ability. However, a large amount of redundant information is generated during feature mapping and enhancement in BLS. In addition, BLS has limited performance in eliminating the negative effects of discrete clustered points in noisy data. To address the above problems, a new robust projection distributed broad learning (RPDBL) is proposed in this article, which finds two projection directions in the feature space to achieve a good separation of projection samples and filter the noise and irrelevant information in the data so as to improve the robustness. Furthermore, to mitigate the problem of discrete clustered points, an additional regularization term is designed to ensure that the optimization problem is positive definite. The experimental results of rolling bearings show that compared with other breadth-based methods, RPDBL exhibits significant advantages in terms of accuracy, kappa, F-score, and precision.