A multi-scale kernel-based network with improved attention mechanism for rotating machinery fault diagnosis under noisy environments

Abstract

Considering that the vibration signals are easily contaminated by strong and highly nonstationary noise, extracting more sensitive and effective features from the noised vibration signals is still a great challenge for intelligent fault diagnosis of rotating machinery. This paper proposes a multi-scale kernel-based network with an improved attention mechanism (IA-MKNet) to overcome this challenge. In the proposed method, an improved attention mechanism (IAM) for multi-scale convolution is firstly developed to adaptively extract the meaningful fault features and automatically suppress noise. Then, due to the inherent multiple time characteristics of vibration signals, an adaptive multi-scale kernel-based residual block with IAM is designed to capture fault features in multi-time scales of vibration signals. Finally, a combination strategy based on an adaptive ensemble learner is proposed to increase the diversity of features by fusing the outputs of multiple IA-MKNets, which further improves diagnostics accuracy and stability. The experimental results, verified by two bearing data sets with noise interference, confirm that the proposed method improves the fault diagnosis accuracy of rotating machinery under noisy environments, whose performance is superior to the other five benchmark methods.