Research Article

Gearbox fault diagnosis convolutional neural networks with multi-head attention mechanism

¹ Zhongyuan University of Technology, Zhengzhou, Henan 451191, China
² Key Laboratory of Optical Sensing and Testing Technology for Mechanical Industry, Zhongyuan University of Technology, 41 Zhongyuan Middle Road, Zhengzhou, Henan Province, China
³ Zhejiang Xiasha Precision Manufacturing Co., Ltd, 389 Rongji Road, Ningbo City, Zhejiang Province, China

^* Corresponding author: xuhangzzti@126.com

Received: 8 December 2023
Accepted: 6 February 2025

Abstract

With the development of intelligent manufacturing systems, data-driven fault diagnosis has become a hot research topic. Traditional data-driven fault diagnosis methods often rely on expert-extracted features, wherein feature extraction process requires considerable effort and affects the final results to a great extent. However, end-to-end fault diagnosis methods based on deep learning can automatically learn feature representations from raw data. In this study, first, the raw vibration signals of various fault states of a planetary gearbox were segmented and preprocessed into input table data types. Second, a convolutional neural network with wide kernels in the first layer was used to extract gear fault features from the raw vibration data of the gearbox. Then, the multi-head attention mechanism was incorporated to focus on different feature spaces and obtain diverse feature information. Finally, using the Softmax layer, the fault features were classified and fault diagnosis of the gearbox was achieved. Validation experiments and comparative analysis indicated that the proposed fault diagnosis model exhibits stronger learning ability as well as a simpler and convenient diagnostic process compared with the traditional methods. The proposed model has broad application prospects in data-driven fault diagnosis.