Unsupervised Learning Model of Sparse Filtering Enhanced Using Wasserstein Distance for Intelligent Fault Diagnosis

Background Deep learning-based fault diagnosis techniques are promising approaches that can eliminate the need for advanced skills in signal processing and diagnostic expertise. Purpose The sparse filtering method is an unsupervised learning method whose parameters play a vital role in obtaining more accurate and reliable results. Thus, their appropriate selection is necessary to obtain more accurate diagnosis of faults in rotating machinery by adaptively collecting the features of the vibration signal. Method In this paper, a novel unsupervised learning method called general normalized sparse filtering (GNSF) based on Wasserstein distance with maximum mean discrepancy (MMD) has been proposed for fault diagnosis. Initially, the objective function of generalized l(r-p/q) norm is optimized to compute the feature sparsity that enhances the regularization performance of the sparse filtering. Whereas, clustering of features is done by Wasserstein distance with MMD highlights the contribution of different features in fault clustering. Conclusion Using GNSF based on Wasserstein distance with MMD, a new intelligent fault diagnosis method is planned and applied to the centrifugal pump and Pelton wheel vibration data under harsh/defect-induced operating conditions. Even with a lesser number of training samples, the proposed approach can successfully identify the various health conditions of the centrifugal pump and Pelton wheel. For instance, at 5% of training samples, the proposed fault identification scheme obtained 99.95% and 99.91% of accuracy. The results confirm the capabilities and effectiveness of the proposed method in terms of efficiency and computation time even for fewer training samples.