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

Area of Science:

• Engineering
• Computer Science
• Machine Learning

Background:

• Data imbalance significantly impacts rotating machinery fault diagnosis using traditional data-driven methods.
• Existing methods struggle with skewed datasets, leading to reduced diagnostic accuracy.

Purpose of the Study:

• To propose a novel Deep Balanced Cascade Forest (DBCF) method for robust fault diagnosis in imbalanced datasets.
• To enhance classification performance by integrating resampling and advanced algorithm design.

Main Methods:

• Developed a multi-channel cascade forest where each channel generates adaptive deep structures trained on independent data.
• Introduced Up-down Sampling, a hybrid method for data rebalancing.
• Designed a novel balanced forest classifier incorporating improved balanced information entropy for attribute selection.

Main Results:

• The DBCF model demonstrated superior stability and effectiveness in handling imbalanced fault diagnosis datasets.
• Comparative experiments showed DBCF outperforms popular deep learning methods in imbalanced classification tasks.
• The synergy between Up-down Sampling and the balanced forest classifier is key to the model's success.

Conclusions:

• The proposed Deep Balanced Cascade Forest offers a powerful solution for rotating machinery fault diagnosis with imbalanced data.
• The fusion of data-level (resampling) and algorithm-level (balanced forest) methods provides significant performance gains.
• DBCF presents a more stable and effective alternative to current deep learning approaches for imbalanced fault diagnosis.