融合时频特征与残差密集网络的医用离心机故障检测方法

Fault Detection Method for Medical Centrifuges Based on Time–Frequency Feature Fusion and a Residual Dense NetworkGong HongYun12、Wu Juan2、Xue Ni2、Cao Yue2、Shang ZiTian3、Zong Hua12*

摘要:

由于医用离心机早期故障特征微弱、多故障耦合及在线预警不足,因此提出一种融合时频特征与残差密集网络的故障检测方法。利用经验模态分解改进伪魏格纳-维尔分布,抑制交叉项并构建精细时频图谱；然后设计融合局部二进制编码的混合卷积模块,并结合密集连接、残差跨层连接和自适应遗传算法,实现故障特征增强与网络参数优化；并依据动态阈值与分类置信度构建分级预警机制。实验结果表明,该方法对正常、轴承磨损、转子失衡及复合故障均具有较高识别精度,在0～20 dB噪声条件下表现出良好的稳定性与抗干扰能力,单样本检测延迟为5.92 ms,并可对渐进性故障实现提前预警。该方法可为医用离心机智能运维与安全管理提供技术支撑。

Abstract:

To address the weak characteristics of incipient faults, the coupling of multiple faults, and the limited online early-warning capability of medical centrifuges, a fault detection method integrating time–frequency features with a residual dense network is proposed. First, empirical mode decomposition is combined with the pseudo Wigner–Ville distribution to suppress cross-term interference and construct refined time–frequency representations. A hybrid convolution module incorporating local binary coding is then designed. Dense connections, cross-layer residual connections, and an adaptive genetic algorithm are further introduced to enhance fault features and optimize network parameters. Finally, a hierarchical early-warning mechanism is established by combining dynamic thresholds with classification confidence. Experimental results show that the proposed method achieves high recognition accuracy for normal conditions, bearing wear, rotor imbalance, and compound faults. It maintains favorable stability and noise robustness over the signal-to-noise ratio range of 0–20 dB, with a detection latency of 5.92 ms per sample, and enables advance warning of progressive faults. The method provides technical support for intelligent maintenance and safety management of medical equipment.