Multi-scale and multi-layer perceptron hybrid method for bearings fault diagnosis

The progressive growth in demand and requirements for bearing problem diagnostics in the operating segment of trains has resulted from an increase in train speed and the development of intelligent operation and maintenance techniques. A multi-scale multi-layer perceptron (MSMLP) hybrid bearing fault diagnosis based on complementary ensemble empirical mode decomposition (CEEMD) is proposed, inspired by the successful application of deep networks in the field of computer vision, combined with the characteristics of large noise and small samples of train bearing data. The original signal is decomposed using CEEMD, the intrinsic mode functions (IMFs) with the highest correlation being retrieved, and the one-dimensional signal can be transformed into a two-dimensional picture using recurrence plots (RPs), which is then separated into multi-channel input to the MSMLP model. The suggested MSMLP model is a hybrid MLP (HMLP) layer and an MLP-based MLP network. The former is used to combine spatial and positional data, while the latter is used to classify data. When the aforesaid approach was compared to classical models for time series classification, it was proven that the method can not only properly identify bearing problems, but also has an accuracy exceeding 99% and a higher capacity to converge than other models.