Multi-channel sensor fusion for real-time bearing fault diagnosis by frequency-domain multilinear principal component analysis

Real-time health condition monitoring of bearings plays a significant role in the functionality of the rotary machinery. Multi-channel sensor fusion can be more robust for identifying diverse bearing fault diagnosis scenarios. However, the high-dimensional data and complex fault scenarios that can occur in the system pose significant challenges for effective fault diagnosis. State-of-the-art artificial intelligence-based bearing fault diagnosis system involves multi-channel sensor fusion, which usually leverages time-frequency analysis, feature extraction, and supervised learning. Nevertheless, those methods usually require a large training dataset for the machine learning model development. This paper proposes a new multi-channel sensor fusion methodology, named frequency-domain multilinear principal component analysis (FDMPCA), by integrating acoustics and vibration signals with different sampling rates and limited training data. Frequency analysis is firstly leveraged to transform the original signals from time to frequency domain, and the frequency responses of heterogeneous channels form a tensor structure named the frequency-domain (FD) tensor. Subsequently, the FD tensor is decomposed by multilinear principal component analysis (MPCA), resulting in low-dimensional process features for fault diagnosis. Finally, the extracted features can be used to train a Neural Network (NN) model for fault diagnosis. To validate the effectiveness of the proposed method, the bearing fault experiments were conducted on a machinery fault simulator while multiple vibration and acoustic signals were collected. Experimental results demonstrated that the proposed approach can effectively identify the machine fault conditions and outperform the benchmark methods given the limited training data.