A Multi-Branch Parallel-Fusion Method for Fault Diagnosis of Shearer Bearings

Abstract

Conventional single-network models struggle to simultaneously capture local transient impacts, temporal dependencies, and contextual correlations from rolling bearing vibration signals. Moreover, traditional serial hybrid architectures suffer from feature dilution, wherein weak fault features extracted by upstream layers are progressively weakened during forward propagation. To address these limitations, this paper proposes CLTrans, a parallel multi-branch fault diagnosis model integrating a one-dimensional convolutional neural network (1D CNN), a two-layer long short-term memory (LSTM) network, and an enhanced Transformer to extract local, temporal, and global features in parallel. The Transformer branch incorporates Rotary Position Embedding (RoPE) to enhance relative positional awareness of periodic impact patterns, and employs Local Window Attention (LWA) to restrict self-attention computation within overlapping sliding windows, thereby reducing computational complexity from quadratic to linear order while suppressing background noise interference. Features from all three branches are concatenated into a unified joint representation for fault classification. Experiments on the Case Western Reserve University (CWRU) dataset for 10-category bearing fault classification yield a macro-averaged accuracy of 99.57%, precision of 99.55%, recall of 99.29%, and F1-score of 99.40%. Compared with the best-performing baseline, Transformer-BiLSTM, CLTrans improves accuracy by 1.29 percentage points and F1-score by 1.76 percentage points. Ablation studies confirm that the three branches form complementary feature representations. The proposed model advances multi-class fault recognition performance on public bearing datasets and provides a methodological foundation for condition monitoring of shearer cutter bearings under noisy industrial environments.