TCCL: template convolution contrastive learning for unsupervised deep clustering of rolling bearing faults

Abstract Unsupervised fault diagnosis of rolling bearings remains a critical challenge in modern industry due to the scarcity of labeled data and the high cost of expert annotation. Existing unsupervised methods, including reconstruction-based approaches (e.g. Autoencoders, VAEs, generative adversarial networks) and contrastive learning (CL) frameworks (e.g. SimCLR, MoCo), struggle with fine-grained fault classification due to shortcut learning and mode collapse problems. These methods often fail to capture the discriminative morphological patterns that distinguish different fault types. To address these limitations, this paper proposes template convolution CL (TCCL), a novel self-supervised framework that integrates learnable template matching with CL. TCCL employs a bank of dynamic templates that convolve with input vibration signals to generate template-response features, explicitly encoding fault-specific waveform structures into the latent space. By maximizing the similarity between signals and their best-matching templates while maintaining contrastive separation between different fault categories, TCCL creates a cluster-friendly representation space. Comprehensive experiments on three public datasets (Case Western Reserve University, Southeast University, and Machinery Failure Prevention Technology) with 5-fold cross-validation demonstrate that TCCL achieves the highest clustering accuracy on all three benchmarks (96.3%, 75.7%, and 80.4%, respectively), consistently outperforming 13 baseline methods including recent 2024–2025 approaches. Moreover, TCCL exhibits strong robustness to the choice of data augmentation strategy, maintaining competitive performance even under minimal augmentation—a property that is particularly valuable for deployment in new industrial domains where optimal augmentation designs are unavailable. The proposed method effectively bridges the gap between unsupervised anomaly detection and fine-grained multi-class fault clustering, providing a promising direction for intelligent maintenance in real-world industrial scenarios.