Bearing Fault Diagnosis Based on Time Series Isolation Slicing and Multi-domain Feature Fusion

Abstract

To address the potential data leakage associated with overlapping slicing and the limited generalization capability of rolling bearing fault diagnosis models under strong noise, small-sample, and class-imbalanced conditions, a parallel CNN–BiLSTM–CBAM method based on time-isolated slicing and multi-domain feature fusion is proposed. A time-isolated data construction strategy is adopted, in which continuous vibration records are first partitioned into independent temporal segments and then sliced separately, thereby reducing the risk of cross-subset information leakage caused by overlapping windows. In addition, −2 dB noise and class-imbalance processing are introduced to simulate complex operating conditions. A dual-branch EWT–CWT mechanism is constructed to collaboratively extract multi-domain time–frequency features, and a gated adaptive fusion module is designed to dynamically estimate channel weights, thereby enhancing fault-sensitive features while suppressing noise interference. A parallel network is employed to capture local impulsive characteristics and bidirectional temporal dependencies, while CBAM further refines critical fault information. The parameter sizes and training times of different models are also preliminarily compared. The proposed method is evaluated on a self-developed bearing test rig under multiple coupled speed–load conditions and on the public CWRU dataset. Experimental results indicate that overlap-first sampling followed by random partitioning may lead to overestimated test accuracy, whereas time-isolated partitioning provides a more reliable indication of model performance on temporally independent test data.

Conflict of Interest Statement

The authors declare no conflicts of interest.