Deep balanced cascade forest: An novel fault diagnosis method for data imbalance

Data imbalance is a common problem in rotating machinery fault diagnosis. Traditional data-driven diagnosis methods, which learn fault features based on balance dataset, would be significantly affected by imbalanced data. In this paper, a novel imbalanced data related fault diagnosis method named deep balanced cascade forest is proposed to solve this problem. Deep balanced cascade forest is a multi-channel cascade forest, in which, each of its channels adaptively generates deep cascade structure and is trained on independent data. To enhance the performance of imbalance classification, the deep balanced cascade forest is promoted from both aspects of resampling and algorithm design. A hybrid sampling method, namely Up-down Sampling, is proposed to provide rebalanced data for each cascade forest channel. Meanwhile, a new type of balanced forest with an improved balanced information entropy for attribute selection is designed as the basic classifier of cascade forest. The good synergy of these two methods is the key to the deep balanced cascade forest model. This good synergy makes deep balanced cascade forest achieve the fusion of data-level methods and algorithm-level methods. Comparative experiments on sufficient imbalanced datasets have been designed to verify the performance of the proposed model, and results confirm that deep balanced cascade forest is much more stable and effective in handling imbalance fault diagnosis problem compared to the popular deep learning methods.