摘要详情

ID / 提交时间

14 / 2025-04-13 21:21:44

标题

Double-Path Residual Network with Attention Mechanism and Bidirectional Gated Recurrent Unit and its Application in Planetary Gear Fault Classification

关键字

residual neural network,attention mechanism,bidirectional gated recurrent unit,fault diagnosis

主题及专题

4、System Monitoring, Fault Diagnosis and Prediction

状态

全文待审

作者

Zhilin Dong / Beijing unversity of technology

Dezun Zhao / Beijing University of Technology

Dongdong Liu / Beijing University of Technology

Honglin Sun / Ltd；World Transmission Technology(Tianjin)Co.

Lingli Cui / Laboratory of Advanced Manufacturing Technology Beijing University of Technology

摘要

The convolution neural network structure in double-path convolution with attention mechanism (AM) and bidirectional gated recurrent unit (BGRU), termed DCAM-BGRU, suffers from the problem of gradient vanishing or exploding with the network increases, affecting the recognition accuracy and stability in the fault classification task. To alleviate this problem, a double-path residual network with an attention mechanism and bidirectional gated recurrent unit (DRAM-BGRU) is developed. First, two parallel one-dimensional residual networks (1D-ResNet) with the optimized structure are designed to extract spatial features and eliminate gradient vanishing or exploding. Second, an AM is designed to catch outstanding impact features for fault samples. Next, a bidirectional gated recurrent unit is set back of 1D-ResNet and AM to mine the temporal features of the sample signal. Finally, the information processed by the hybrid module is summarized through global average pooling, and the fully connected layer is used to output the final classification result. The developed technique is tested using one set of planetary gear data and achieves a satisfied classification accuracy of 100%. Compared with the other five advanced methods, the proposed technique can detect planetary gear faults much better.

The convolution neural network structure in double-path convolution with attention mechanism (AM) and bidirectional gated recurrent unit (BGRU), termed DCAM-BGRU, suffers from the problem of gradient vanishing or exploding with the network increases, affecting the recognition accuracy and stability in the fault classification task. To alleviate this problem, a double-path residual network with an attention mechanism and bidirectional gated recurrent unit (DRAM-BGRU) is developed. First, two parallel one-dimensional residual networks (1D-ResNet) with the optimized structure are designed to extract spatial features and eliminate gradient vanishing or exploding. Second, an AM is designed to catch outstanding impact features for fault samples. Next, a bidirectional gated recurrent unit is set back of 1D-ResNet and AM to mine the temporal features of the sample signal. Finally, the information processed by the hybrid module is summarized through global average pooling, and the fully connected layer is used to output the final classification result. The developed technique is tested using one set of planetary gear data and achieves a satisfied classification accuracy of 100%. Compared with the other five advanced methods, the proposed technique can detect planetary gear faults much better.