Pelton Wheel Bucket Fault Diagnosis Using Improved Shannon Entropy and Expectation Maximization Principal Component Analysis

Background Pelton wheel works on Newton's law which converts the kinetic energy of fluid into mechanical energy. Bearing, nozzle, servomotor and buckets are the main components of the Pelton wheel that are prone to defects. Corrosion by reactive materials, degradation by strong slurry particles, the involvement of some metallurgical defects, cavitation, and poor bearing lubrication are some of the causes which induce defects in the Pelton wheel. These failures result in significant turbine disruption, costly disassembly, and, in some cases, full Pelton wheel shutdown. Hence, it becomes a necessity to monitor the Pelton wheel through some suitable methods. Purpose A novel artificial intelligence-based method has been investigated to describe the health condition of a Pelton wheel. Traditionally, extracted features from stationary wavelet transform (SWT) decomposed signal to increase the complexity and affect the classification accuracy. This issue is resolved by developing a new fault diagnosis scheme using improved Shannon entropy based on expectation maximization principal component analysis (EM-PCA) and extreme learning machine (ELM). Methods In the proposed scheme, F-score is initially applied to select features and construct the feature matrix. At the same time, EM-PCA is used to reduce the dimension of the constructed feature matrix, which reduces the correlation between data and eliminate the redundancy to retain the essential features for the ELM classification model. Conclusion The effectiveness of the proposed scheme is compared with other reduction techniques used for the purpose. A comparison has also been made with other classification methods. The results show that EM-PCA with improved Shannon entropy can effectively eliminate correlation and redundancy of data. Further, the use of the ELM can take on better adaptability, faster computation speed and higher classification rate. The proposed method is fast as it takes 0.0020 s of computation time for both training and testing with 89.14% and 96.33% training and testing accuracies, respectively.