Deep forest based intelligent fault diagnosis of hydraulic turbine

Abstract

Deep neural networks (DNNs) for intelligent machinery fault diagnosis require a large amount of training data, powerful computational facilities and have many hyper-parameters that have to be carefully tuned to ensure maximum performance. Deep forest, as a novel alternative to the deep learning framework, has the potential to overcome these shortcomings. In this study, a deep forest-based end-to-end intelligent fault diagnosis method is proposed for hydraulic turbine, in which multi-grained scanning is first used to transform fault feature representations from raw data and enhance fault feature learning ability, and then cascade structure is constructed with different types of random forests to learn fault features level by level and classify faults. The effectiveness of the proposed method is validated using the experimental dataset under twelve conditions, and its practicability is validated using a simulated dataset generated by adding white Gaussian noise to raw experimental signals. The results show that the proposed method is able to adaptively mine available fault features from measured signals, and its diagnosis accuracy is better than that obtained by existing methods. More importantly, the proposed method has better robustness to noise and is less limited to the number of training data.