Fault diagnosis of RV reducer based on denoising time–frequency attention neural network

• Published in: Expert systems with applications 2024-03, Vol.238, p.121762, Article 121762
• Main Authors: Jiang, Kuosheng, Zhang, Chengsong, Wei, Baoliang, Li, Zhixiong, Kochan, Orest
• Format: Article
• Language: English
• Subjects: Angular contact ball bearings; Attention module; DNCNN; InceptionV4; RV reducer
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2023.121762

RV reducer operating conditions are complex, often with low speed, reciprocating, non-integral cycle and other characteristics, coupled with the RV reducer itself complex structure and fault signal instantaneous, so the fault signal sample is small, which brings challenges to the bearing fault diagnosis of RV reducer. Therefore, an improved DNCNN denoising algorithm is proposed to realize the failure mode recognition of RV reducer. The algorithm improves the basic unit module (Conv(3*3) + BN + Relu) in DNCNN network, which can effectively solve the problem that the stacking of convolutional blocks leads to a large number of network parameters, makes the training time long, and most of the training parameters are concentrated in a certain layer of the network. The SE attention module is introduced to avoid the problem that the CNN model cannot effectively highlight the key points when extracting features. At the same time, the Swish activation function is used to replace the Relu activation function, which improves the phenomenon that the latter cannot learn the value of the function input less than 0, and the parameters cannot be updated. Based on improving the DNCNN denoising algorithm, the neural network fault diagnosis method of the main bearing state of the three RV reducers after denoising is studied, and the CBAM attention mechanism module is added to the network to obtain the InceptionV4-CBAM network model. Multiple sets of experimental analysis show that the failure mode recognition rate of the main bearing of the RV reducer under noise interference by the Improved DNCNN combined with InceptionV4-CBAM is 97.3 %.