军事知识服务平台
基于CNN-BCA-ISA的电机故障诊断研究
Research on Motor Fault Diagnosis Based on CNN-BCA-ISA
摘要 电机设备在工业生产领域扮演着至关重要的角色,但长期运行及复杂多变的内外部条件使其面临诸多故障挑战.鉴于此深入探讨了应用卷积神经网络技术来强化电机故障诊断效能的前沿领域,研究工作聚焦于生产线潜在的电机故障场景,提出基于BCA-ISA协同优化的CNN架构,通过三级递进式设计突破传统诊断瓶颈:首先整合面向一维时序数据的多尺度窗口平移-叠加噪声的复合增强策略,在相同数据集条件下可提升模型泛化能力达10%;其次针对模型参数固化问题,创新性融合蜂群优化算法(BCA)的全局搜索特性与CNN的局部特征提取能力,构建参数空间自适应映射机制,实现复杂工况下模型参数的自优化配置,相较于传统梯度下降方法参数调整速率提升16.2%;最后,为强化特征捕获能力,设计大卷积核-瓶颈层-改进自注意力机制(ISA)的协同结构:大卷积核+瓶颈层通过扩大感受野与特征压缩技术,解决了高频特征捕获效率问题,ISA模块引入位置权重与梯度门控机制,突破时序信号长程依赖建模瓶颈,实验验证该结构对故障高频特征的识别效率较基线模型提升31%.
Abstract:
Motors play a critical role in industrial production,yet their long-term operation under complex and variable conditions poses significant fault challenges.To address this,a convolutional neural network(CNN)architecture with BCA-ISA collaborative optimization was proposed,featuring a three-stage progressive design:firstly,a composite enhancement strategy integrating multi-scale window shifting and noise superposition for 1D time-series data was implemented,improving model generalization capability by 10%under identical dataset conditions.Secondly,targeting parameter solidification issues,innovatively combined the global search characteristics of the bee colony algorithm(BCA)with CNN's local feature extraction capabilities.This established an adaptive parameter-space mapping mechanism enabling self-optimizing configuration under complex operating conditions,achieving a 16.2%acceleration in parameter adjustment compared to traditional gradient descent methods.Finally,a collaborative structure integrating large kernels,bottleneck layers,and improved self-attention(ISA)mechanisms was designed:the large kernel+bottleneck layer combination resolved high-frequency feature capture inefficiencies through expanded receptive fields and feature compression.The ISA module overcomed long-range dependency modeling bottlenecks via positional weighting and gradient gating mechanisms experimental validation confirms a 31%enhancement in high-frequency fault feature recognition versus baseline models.
刊 名 电气传动 ISTIC
年,卷(期): 2026, 56(10)
分类号: TP278
机标分类号: TP391.41 TM743 TP18
在线出版日期: 2026年10月28日
基金项目: QNJJ-PY-2024-04:中国机械工业集团有限公司青年科技基金,