方位(导航)
断层(地质)
振动
信号(编程语言)
特征(语言学)
融合
计算机科学
声学
信息融合
模式识别(心理学)
语音识别
人工智能
地质学
物理
地震学
哲学
程序设计语言
语言学
作者
Aijun Hu,Kaijie Yang,Bowen Yang,Zhuohao Zhou,Xianze Li,Ling Xiang
标识
DOI:10.1088/2631-8695/ade849
摘要
Abstract Multisource mechanical signals provide reliable and robust information about machinery operating status. However, the effective integration of such heterogeneous information remains a significant challenge in intelligent fault diagnosis. While each sensor modality captures both shared and modality-specific features, fusing the consistent information across signals is crucial for improving diagnostic performance. In this paper, a consistent feature fusion network (CFFN) is proposed, with a particular focus on the correlation-guided integration of vibration and acoustic signals. Two parallel enhanced convolutional neural networks are employed to extract representative features from raw multisource data. To explicitly exploit the latent correlations between modalities, a canonical correlation analysis (CCA)-based fusion module is introduced to extract and integrate the consistent feature components. The effectiveness of the proposed fusion strategy is validated through extensive experiments. The proposed method achieves a classification accuracy of 99.52% on the NU205E dataset, and consistently exceeds 99% across other datasets. Under high noise conditions, the model maintains robust performance with an accuracy of 92.97%. Furthermore, ablation studies and quantitative clustering evaluations confirm the advantages of the proposed consistent feature fusion mechanism.
科研通智能强力驱动
Strongly Powered by AbleSci AI