A CEEMDAN-GAF and Cross-Attention-Based Multimodal Fault Diagnosis Method for Ball Screws

作者：Youjin Cheng, Ping Wang, Xuan Dong, Hongjin Chen · 发表于：Chinese Control and Decision Conference · 年份：2026 · DOI：10.1109/ccdc69976.2026.11560241

Traditional single-modal fault diagnosis methods for ball screws inevitably suffer from the loss of multi-physics coupling information, leading to degraded diagnostic accuracy and limited robustness. To overcome this limitation, a multimodal deep learning framework, termed the CEEMDAN-GAF CrossAttention Network (CGCAN), is proposed for ball screw fault diagnosis. The framework integrates adaptive decomposition of vibration and acoustic signals using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), temperature signal encoding using Gramian Angular Fields (GAF) and dynamic multimodal feature fusion based on a crossattention mechanism. By exploiting complementary information across vibration, acoustic and temperature modalities, CGCAN achieves robust representation of fault-related spatiotemporal patterns. Experimental results on a self-built ball screw dataset demonstrate that CGCAN consistently outperforms TCN, LSTM and MCNN-LSTM baselines in terms of accuracy, MacroF1, precision and recall. Ablation studies further confirm the effectiveness of each key component.