An Intelligent Classification Approach via Multiple Classifier Fusion and Its Application to the Fault Diagnosis

The performance of classifiers plays a crucial role in identifying fault categories under mechanical fault diagnosis task. This paper presents a new intelligent classification approach via NA-MEMD and multiple classifier fusion for fault diagnosis, which comprises four stages. First, NA-MEMD extracts time-domain and frequency-domain features from the original vibration signal. Second, Sensitive feature extraction using improved Fisher’s criterion method. Third, the sensitive feature sets (SFS) are input to Depth Gaussian Restricted Boltzmann Machine (DG-RBM), RNN, and CNN classification methods to attain the complementary benefits and substantial fusion of various classifiers. Final, the Bayesian belief approach is utilized for fusing the classification results of multiple classifiers to obtain the diagnosis results. Experiments on rolling bearing datasets reveal that the presented approach can precisely detect the fault conditions and provide a classification efficiency superior to the single classifiers and ensemble of all classifiers. The experimental results have revealed the efficiency, generalization, and robustness of the multiple classifier fusion.