The application of audio signals in gear fault diagnosis based on deep learning methods: an end-to-end ap-proach
پذیرفته شده برای ارائه شفاهی ، صفحه 1-12 (12) XML اصل مقاله (1.88 MB)
نویسندگان
1دانشکده مهندسی مکانیک، دانشگاه صنعتی امیرکبیر
2Acoustics Research Lab., Mechanical Engineering Department, Amirkabir University of Technology
چکیده
Diagnosing gearbox faults based on audio signal has received less attention in researches, alt-hough due to the non-contact nature of the microphone, it makes the diagnosis process more accessible. In this article, based on the methods of deep learning, the diagnosis of crack and uniform wear of the gearbox in three different scenarios of (1) constant fault severity and working conditions, (2) constant fault severity and different working conditions, (3) different fault severity and different working conditions have been investigated. State-of-the-art methods of Convolutional Neural Network (CNN), Deep Residual Neural Network (DRN) and a proposed hybrid network of CNN and Long Short-Term Memory (LSTM), all applied based on end-to-end approach, have been investigated. The results show that the CNN+LSTM has a better performance than other methods, in such a way that in the most difficult case, i.e. different fault severity and different working conditions, it classifies the faults with an accuracy of 88.8%. In addition, the computational cost of training the pro-posed network is less than other networks.
کلیدواژه ها
موضوعات
 
Title
The application of audio signals in gear fault diagnosis based on deep learning methods: an end-to-end ap-proach
Authors
Hassan Alavi, Abdolreza Ohadi
Abstract
Diagnosing gearbox faults based on audio signal has received less attention in researches, alt-hough due to the non-contact nature of the microphone, it makes the diagnosis process more accessible. In this article, based on the methods of deep learning, the diagnosis of crack and uniform wear of the gearbox in three different scenarios of (1) constant fault severity and working conditions, (2) constant fault severity and different working conditions, (3) different fault severity and different working conditions have been investigated. State-of-the-art methods of Convolutional Neural Network (CNN), Deep Residual Neural Network (DRN) and a proposed hybrid network of CNN and Long Short-Term Memory (LSTM), all applied based on end-to-end approach, have been investigated. The results show that the CNN+LSTM has a better performance than other methods, in such a way that in the most difficult case, i.e. different fault severity and different working conditions, it classifies the faults with an accuracy of 88.8%. In addition, the computational cost of training the pro-posed network is less than other networks.
Keywords
Fault diagnosis, gearbox, Deep learning, End-to-end approach