文献类型：期刊文献

英文题名：Folding Convolutional Neural Network for Rotating Machinery Fault Diagnosis

作者：Lu, Tiantian[1];Zhu, Hongbin[2];Dai, Jisheng[3];Lai, Huadong[4];Xu, Weichao[1]

机构：[1]Guangdong Univ Technol, Sch Automat, Guangzhou 510006, Peoples R China;[2]Guangdong Polytech Normal Univ, Sch Elect & Informat, Guangzhou 510665, Peoples R China;[3]Donghua Univ, Coll Informat Sci & Technol, Shanghai 201620, Peoples R China;[4]Guangdong Ocean Univ, Sch Elect & Informat Engn, Zhanjiang 524088, Peoples R China

年份：2026

外文期刊名：IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

收录：SCI-EXPANDED(收录号：WOS:001767452900001)、、EI(收录号：20262020741022)、Scopus(收录号：2-s2.0-105039240636)、WOS

基金：This work was supported in part by the National Natural Science Foundation of China under Grant 62571146 and Grant 62501175, in part by Guangdong Science and Technology Department under Grant 2024A1515011803, Grant 2026A1515011625, and Grant 2026A1515011437, and in part by Natural Science Foundation of Shanghai Basic Research Funding under Grant 25ZR1401002. Paper no. TII-26-1987.

语种：英文

外文关键词：Fault diagnosis; Educational institutions; Modeling; Convolutional neural networks; Long short term memory; Transfer learning; Kernel; Machinery; Transformers; Convolution; folding convolutional neural network; neural network; rotating machinery; transformer

外文摘要：One-dimensional convolutional neural network (1D-CNN)-based methods are widely used in fault diagnosis for rotating machinery. However, traditional 1D-CNN-based approaches are limited by the fixed receptive field of their convolutional kernels, which restricts their ability to capture long-range dependencies in signal data. To address the issue, a folding convolutional neural network (FCNN) is proposed. By folding the input signals before convolution, the FCNN enables a relative shift between distant information and the CNN's receptive field, allowing the model to effectively utilize both local and long-range features. Building on the FCNN, a multiview fault diagnosis framework is constructed through the integration of a transformer encoder and a long short-term memory (LSTM) network. Transfer learning is employed to optimize the initial parameters of the model. The proposed method first uses a three-layer FCNN to extract features from each channel individually, followed by feature fusion. The fused features are then analyzed by a transformer encoder, which captures global contextual relationships through self-attention, and further processed by an LSTM to model temporal dynamics. Experiments conducted on one private and three public datasets demonstrate the effectiveness, robustness, noise tolerance, and advanced performance of the proposed model.