Multiscale Deep Subspace Clustering Network With Hierarchical Fusion Mechanism for Mechanical Fault Diagnosis

Deep learning-based unsupervised fault diagnosis methods have shown excellent performance and garnered significant research interest. However, they encounter two significant challenges: the separate optimization of deep fault feature learning and fault pattern recognition, and the multiscale characteristic in signals under nonstationary operating conditions. In this article, a novel multiscale deep subspace clustering network with hierarchical fusion mechanism (MSDSC-HFM), is proposed for unsupervised mechanical fault diagnosis. First, convolutional neural networks with multiple kernel sizes are employed to learn multiscale deep fault features, and the self-expressive layer is embedded to obtain the scale-specific coefficient matrix. Second, a hierarchical fusion mechanism is designed to capture the fused deep fault feature with complementary information. Specifically, two custom-designed fusion blocks are developed to realize interactive reinforcement between high-level semantic fault features and low-level detail fault features. Finally, an unsupervised fault diagnosis loss considering consistency and diversity constraints is proposed to guide the learning of the common affinity matrix and spectral clustering is performed to achieve efficient and accurate fault pattern recognition. To validate the effectiveness of the proposed method, extensive experiments were performed on real-world datasets. The experimental results demonstrate that the MSDSC-HFM achieves superior unsupervised fault diagnosis performance under nonstationary operating conditions.