Multi-Scale Attention Entropy for Robust Fault Diagnosis and Fault Severity Estimation in Rotating Machines Without Prior Knowledge

Assessing the severity levels of faults in rotating machines is a critical endeavour within the industry, owing to the challenging nature of the noisy working environment and the subtle fault characteristics present in the acquired signals. In this study, a new feature extraction method named multi-scale attention entropy (MSAE), which is a combination of the attention entropy (AttnEn) and the multi-scale entropy (MSE) to extract more discriminative features from the signals, is introduced and investigated. A comparison between the MSAE and randomly selected feature vectors built from a set of 32 statistically and probabilistically features, with the same length, is made to show the performance and ability of the MSAE method. The comparison also includes consideration of the feature vector extracted from the multi-scale sample entropy (MSSE), which is the earliest version of the MSE. Subsequently, all ten feature vectors are input into a support vector machine (SVM) classifier for fault diagnosis and estimation of fault severities. Finally, the performance of the methods is compared for two scenarios, fault diagnosis (FD) and fault diagnosis and severity estimation (FD&SE), on two challengeable datasets. The first dataset, the Case Western Reserve University bearing (CWRU) dataset, is a bearing fault dataset, while the second one, the Korea Advanced Institute of Science and Technology (KAIST) dataset, is a rotor-bearing fault dataset. After twenty iterations, the MSAE-SVM model achieved an average FD accuracy of "99.58%±0.57%" for the CWRU dataset and "93.05%±0.66%" for the KAIST dataset. In addition, the FD&SE accuracy of the MSAE-SVM model for CWRU and KAIST datasets were \(\:\text{98.64\%±0.68\%}\) and \(\:\text{95.75\%±0.71\%}\), respectively. According to the accuracy tolerance of the feature vectors results from the MSAE-SVM, which is lower than those of other feature vectors, the presented model is more robust in testing accuracy. The presented model is also free of prior knowledge classification and presents much higher mean accuracy among other models used for comparison.