Modified convolutional neural network with global average pooling for intelligent fault diagnosis of industrial gearbox

Abstract

Gearboxes are key transmission components and widely used in various industrial applications. Due to the possible operational conditions, such as varying rotational speeds, long period of heavy loads, etc., gearboxes may easily be prone to failure. Condition Monitoring (CM) has been proved to be an effective methodology to improve the safety and reliability of gearboxes. Deep learning approaches, nowadays, further enable the CM with more powerful capability to exploit faulty information from massive data and make intelligently diagnostic decisions. However, for most of conventional deep learning models, such as Convolutional Neural Network (CNN), a large amount of labelled training data is a prerequisite, while to obtain the labelled data is usually a laborious and time-consuming job and sometimes even unattainable. In this paper, to handle the case of only a limited labelled data is available, a modified convolutional neural network (MCNN) is proposed by integrating global average pooling (GAP) to reduce the number of trainable parameters and simplify the architecture of deep learning model. The proposed MCNN improves the traditional CNN’s ability in fault diagnosis with limited labelled data. Two experimental gearbox datasets are utilized to demonstrate the effectiveness of the proposed MCNN method. Compared with traditional deep learning approaches, namely LSTM, CNN and its variant methods, the experimental results show that the proposed MCNN with higher discrimination and generalization ability in fault classification and diagnostics under the scenario of limited labelled training samples.