Dual-Ensemble Multi-Feedback Neural Network for Gearbox Fault Diagnosis

Abstract

Gearboxes have been widely used in heavy industries. Thus, accurately diagnosing complex health states of the gearbox under different operating conditions is important. To solve this problem, this study presents a systematic fault diagnosis framework using a dual-ensemble multi-feedback neural network. First, salient fault features are extracted by effectively decomposing, denoising, and reconstructing raw vibration signals with ensemble empirical mode decomposition (EEMD). Then, an output-input-hidden feedback (OIHF) Elman neural network (ENN) is constructed, which integrates multiple feedback from various layers. This multi-feedback neural network is capable of effectively building mapping relationships between features and health states. Finally, an optimized AdaBoost-bagging dual-ensemble algorithm is designed to further boost the performance of the base network. The AdaBoost ensemble can greatly reduce generalization errors, and the bagging ensemble can improve algorithm stability by reducing the influence of singular samples. The proposed framework is validated using two datasets from gearboxes. Experimental results show that the framework enables steady and reliable gearbox fault diagnosis under various operating conditions.