A novel bearing fault diagnosis method with feature selection and manifold embedded domain adaptation

Abstract

Traditional fault diagnosis models assume that the training and test data sets have the same feature distribution, but in practice the distribution between the training and test sets varies considerably, making it difficult to achieve the desired fault diagnosis performance. Thus, a diagnosis method based on feature selection and manifold embedding domain adaptation is proposed in this paper. First, the signal is decomposed by variational modal decomposition to obtain multiple modal components, and the entropy, time domain and frequency domain features of each modal component are extracted to form a mixed feature set. Second, it proposes a feature evaluation index based on Fisher scores and feature domain differences to select features that are conducive to cross-domain fault diagnosis and transfer learning. Then, the geodesic flow core is constructed to learn the transformation feature representation in the Grassmann manifold space to avoid features are distorted. Finally, the domain adaptation is performed by minimizing the discrepancy in the joint probability distribution between the same category and maximizing the discrepancy between the different categories. Based on the results of multi-index experiments, the method in this paper is superior to other existing methods.

Combining transfer learning and hyperspectral imaging to identify bruises of pears across different bruise types

“…Theoretically, the shortest distance between two points in the manifold space should be a geodesic line. By constructing a geodesic line flow Φ(t) (0 ≤ t ≤ 1) between two points for feature mapping, the converted manifold features are expressed as w = g ( x ) = Φ ( t ) T x (Yang & Zheng, 2022). Take any two transformed eigenvectors w i and w j , whose inner product can define a kernel function: $$$$\begin{equation}\left\langle {{{w}_i},{{w}_j}} \right\rangle = \int_{0}^{1}{{{{{(\Phi {{{(t)}}^{\mathrm{T}}}{{x}_i})}}^{\mathrm{T}}}(\Phi {{{(t)}}^T}{{x}_j})dt = }}{{x}_i}^{\mathrm{T}}G{{x}_j}\end{equation}$$$$ where G ϵ R k × k is a positive semidefinite matrix obtained by singular value decomposition.…”

“…The core idea of MEDA is constructing a Geodesic Flow Kernel in Grassmann manifold space. Then extract inherent manifold feature representation related to pear bruise information, and dynamically evaluate the condition distribution and edge distribution of manifold features by defining a cross‐domain adaptive factor ( µ ) (Yang & Zheng, 2022). In the end, based on the principle of structural risk minimization, iterative solution of a cross‐domain classifier ( f ) to predict the labels of samples in the target domain.…”

“…Theoretically, the shortest distance between two points in the manifold space should be a geodesic line. By constructing a geodesic line flow Φ(t) (0 ≤ t ≤ 1) between two points for feature mapping, the converted manifold features are expressed as w = g ( x ) = Φ ( t ) T x (Yang & Zheng, 2022).…”

Combining transfer learning and hyperspectral imaging to identify bruises of pears across different bruise types

“…edge distribution of manifold features by de ning a cross domain adaptive factor (µ) (Songyu & Xiaoxia, 2022). In the end, based on the principle of structural risk minimization, iterative solution of a cross domain classi er (f) to predict the labels of samples in the target domain.…”

Bearing fault diagnosis based on variational autoencoder and non-local block wide kernel convolutional neural network

“…4 To ensure personnel safety and reduce economic losses, the research of bearing fault diagnostics is very important. 5,6 Due to its potent feature extraction and fault classification capabilities, deep learning (DL)-based bearing fault diagnosis technology has seen considerable success over the past 10 years. [7][8][9] The model based on DL can automatically capture the latent feature of data without prior knowledge and expert experience, and it can classify faults effectively and accurately.…”