A two-stage fault diagnosis methodology for rotating machinery combining optimized support vector data description and optimized support vector machine

Most intelligent fault diagnosis methods of rotating machinery generally consider that normal samples and fault samples as equally important for pattern recognition training. It ignores that the rotating machinery is in a normal state most of the time, and the collected normal samples are far more than the fault samples. Considering this situation, this paper proposes a two-stage intelligent fault diagnosis methodology for rotating machinery combing optimized support vector data description (SVDD) and optimized support vector machine (SVM). Specifically, SVDD is applied for fault detection, and SVM is applied for fault identification. The parameters of SVDD and SVM are optimized by the grasshopper optimization algorithm (GOA). Multiscale entropy (ME) is used for feature extraction, which is the input feature vector of SVDD and SVM. Under the framework of the proposed methodology, the strengths and weaknesses of 8 different entropy-based indicators in feature extraction are explored. The analysis results of three cases prove the availability and universality of the proposed methodology. According to the analysis results of three case analyses, the permutation entropy (PE)-based indicators are not promising for fault detection. Refined composite multiscale fuzzy entropy (RCMFE) is recommended for fault detection and fault identification to generalize the universality. Based on RCMFE, the fault detection accuracy in the three cases is more than 99%, and the fault identification accuracy is more than 94%. This methodology, therefore, provides a powerful tool for rotating machinery fault diagnosis, which is beneficial to practical application.