Multisensor-Based Bearing Fault Diagnosis Using Bag-of-Correlated-Time–Frequency Features Under Harsh Industrial Background

In industrial processes, prolonged usage of bearings degrades the perfo-rmance of three-phase induction motors (IMs), which leads to substantial economic losses. In that context, we present an intelligent, early, and robust bearing fault diagnosis model based on wavelet coherence (WC)-driven multiclass support vector machine (SVM) classifier using a feature image retrieval (IR) system incorporating nonlinear features. The bearing vibration and sound signals are collected using two different sensors from a real-time experimental bench, considering several realistic challenges of industrial applications. Segmentation of collected signal samples per rotational speed, up/down sampling, and normalization is done consecutively. Further, coherence between sound and vibration samples is estimated in the time–frequency plane to extract strong features using the bag-of-speeded-up robust feature detector. Then, the SVM classifier fits for training and testing with the optimized feature dimension. Finally, the proposed mutisensor-based WC-IR-SVM method is reported to outperform the conventional 2-D convolutional neural network and bidirectional long short-term memory model even under a harsh background.