融合CNN与注意力机制的医疗影像设备故障溯源与预测维护系统-《中国医学物理学杂志》

文章信息/Info

Title:: Fusion of CNN and attention mechanism for fault traceability and predictive maintenance system of medical imaging equipment

文章编号:: 1005-202X（2026）09-1299-10

作者:: 祝传亮1; 顾增1; 辛亚珍1; 周天任1; 周军华2; 1.东南大学附属中大医院无锡分院设备科，江苏无锡 214011；2.江南大学附属医院医学工程处，江苏无锡 214062

Author(s):: ZHU Chuanliang1; GU Zeng1; XIN Yazhen1; ZHOU Tianren1; ZHOU Junhua2; 1. Equipment Department, Zhongda Hospital Wuxi Branch, Southeast University, Wuxi 214011, China 2. Department of Medical Engineering, Affiliated Hospital of Jiangnan University, Wuxi 214062, China

关键词:: 医疗影像设备; 故障溯源; 预测维护; 卷积神经网络; 注意力机制; 通道注意力; 空间注意力; 时序数据处理

Keywords:: Keywords: medical imaging equipment fault traceability predictive maintenance convolutional neural network attention mechanism channel attention spatial attention time-series data processing

分类号:: R318

DOI:: DOI:10.3969/j.issn.1005-202X.2026.09.019

文献标志码:: A

摘要:: 针对医疗影像设备故障溯源效率低、维护模式被动、诊断精度不足等问题，提出一种融合卷积神经网络（CNN）与双通道注意力机制（SE-Net+CBAM）的智能故障溯源与预测维护模型。首先，构建覆盖CT机、MRI机、DR机3种典型设备的多维度故障数据集，设计“清洗-标准化-时序序列构建-特征筛选”4步预处理流程，解决故障数据缺失、噪声干扰、维度冗余等问题；其次，搭建“特征提取-注意力增强-双任务输出”一体化模型，采用1D-CNN提取时序故障数据的局部与深层特征，通过SE-Net通道注意力筛选故障关键通道、CBAM空间注意力定位故障异常位置，双任务分支分别实现故障类型溯源与故障剩余时间、严重等级预测，采用加权求和总损失函数实现双任务协同优化；最后，通过对比实验、消融实验及鲁棒性、实时性验证，验证模型有效性。实验结果表明，本文模型故障溯源准确率达96.8%，预测误差率降至4.3%，鲁棒性与实时性均满足医疗机构运维需求，相比传统机器学习方法与单一CNN模型性能优势显著，能够实现医疗影像设备故障从“事后维修”向“主动预测维护”转变，为医疗影像设备智能化运维提供可靠技术支撑。

Abstract:: Abstract: An intelligent fault traceability and predictive maintenance model integrating convolutional neural network (CNN) and dual-channel attention mechanism (SE-Net+CBAM) is developed to address the problems of poor fault traceability, passive maintenance mode, and insufficient diagnostic accuracy existing in medical imaging equipment. A multi-dimensional fault dataset covering 3 typical types of medical imaging equipment (CT, MRI, and DR) is firstly constructed, and a 4-step preprocessing process consisting of cleaning, standardization, time-series construction, and feature selection is designed to solve the issues including missing fault data, noise interference, and dimensional redundancy. Subsequently, an integrated model of feature extraction, attention enhancement, and dual-task output is established. This model adopts a 1D-CNN to extract local and deep features from time-series fault data. It uses the SE-Net channel attention to screen critical fault channels, and the CBAM spatial attention to locate abnormal fault regions. Furthermore, dual-task branches are employed to realize fault type traceability and prediction of remaining time to failure and fault severity level. A weighted-sum total loss function is applied to achieve dual-task collaborative optimization. Finally, the effectiveness of the model is verified through comparative experiments, ablation experiments, as well as robustness and real-time performance verification. Experimental results show that the proposed model achieves a fault traceability accuracy of 96.8%, and reduces the prediction error rate to 4.3%. Both robustness and real-time performance met the operation and maintenance requirements of medical institutions. The proposed model outperforms the traditional machine learning methods and single CNN models. It can realize the transformation of medical imaging equipment faults from post-maintenance to active predictive maintenance, and provide reliable technical support for the intelligent operation and maintenance of medical imaging equipment.

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

文章信息/Info

备注/Memo

常用功能

导航/Navigate

工具/Tools

统计/Statistics