Machine learning algorithms for optimization of bearing fault diagnosis in wind systems: a comparative study

The research compares the performance of a pair of machine learning toolkits PCA-ANN and LDA-ANN in optimising the fault diagnosis performance of a turbo machinery, a wind turbine, where bearings have become one of the most unreliable parts in mechanical systems. Our dataset, a 50-days record of a vibration signal acquired on a high-speed shaft of a 2 MW wind turbine, enabled us to obtain 15 statistical parameters in the time domain and resort to Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) to reduce the dimensionality without losing much of the variance (97.6 percent) (PCA) or, at the same time, without diminishing the possibilities to separate the data in different classes (maximize the separability between classes) (LDA). Biometric performance of these reduced features was subsequently trained using a 10 neurons hidden layer in an ANN (PCA-ANN tested with accuracy = 80% loss = 0.059 and LDA-ANN tested with accuracy = 86.7%, loss = 0.031). LDA-ANN has proven to be better, requiring less time to execute (24.36 s vs. 28.39 s) and report fewer errors, which further establishes the strength of LDA in supervised fault detection. These findings present strong evidence of the effectiveness of a combination of statistical feature extraction, dimension reduction, and ANN-based classification in predictive maintenance, and at the same time, they indicate the potential to be advanced using additional datasets or hybrid modeling models.