A Multimodal Progressive Fusion Bearing Fault Diagnosis Algorithm Based on Residual Network

• Published in: IEEE sensors journal Vol. 25; no. 13; pp. 23857 - 23868
• Main Authors: Zhu, Wenbing, Ni, Haibin, Li, Zhuo, Cao, Ji, Ni, Bo, Chang, Jianhua
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.07.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Aerospace engines; Algorithms; Bearing fault diagnosis; Convolution; Convolutional neural networks; Data mining; Data models; Downtime; Effectiveness; Fault diagnosis; Feature extraction; High speed rail; Industrial applications; Industrial Internet of Things; Maintenance costs; multimodal; Nickel; Predictive maintenance; progressive feature fusion; Representations; residual networks; Residual neural networks; Rotating machinery; Time-frequency analysis; Training; Wind turbines
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2025.3571201

Bearings are crucial components of rotating machinery in critical industrial equipment such as wind turbines, high-speed trains, and aerospace engines. Existing methods for bearing fault diagnosis are generally confined to superficial integration of multisensor or multidomain data, constrained by either poor heterogeneous information integration in early fusion approaches or information loss and imbalanced modality representations caused by late fusion strategies, resulting in limited diagnostic effectiveness under complex and dynamic industrial operating conditions. In order to solve this issue, we propose a multimodal progressive fusion bearing fault diagnosis algorithm based on residual networks (MMPro-ResNet). The algorithm integrates multisensor and multidomain data and automatically extracts fault features using residual networks. Then, an improved progressive feature fusion technique is applied to optimize the use of the multimodal features, which aims to allow earlier layers to access later fused features, avoiding the loss of important information and improving the fusion representation over multiple iterations. The diagnostic efficacy of the proposed method is validated using two different bearing datasets, achieving a diagnostic accuracy of 99.97% for composite faults. This advancement shows great potential for implementation on industrial internet of things (IoT) platforms, especially in scenarios, such as power generation and transport, where predictive maintenance is required, reducing unplanned downtime and maintenance costs.