摘要
针对旋转机械声学信号中存在的非平稳性强与噪声干扰显著等问题,以及现有方法在时间-尺度建模能力不足、依赖手工时频变换且模型复杂不利于边缘部署的局限,提出了一种融合多尺度动态注意力与1D-2D卷积结构的轻量级端到端故障诊断模型(Multiscale Dynamic Attention and 1D-2D convolutional Fusion Network,MDAF-Net)。该模型集成4项关键模块:首先,构建多尺度动态加权特征提取(Multiscale Dynamic Weighting Feature Extractor,MDW-FE)模块,结合多尺度卷积核与自适应加权机制,以增强对非平稳声学特征的感知能力;其次,设计多尺度映射层(Reshaped Multiscale Projection,RMP),实现一维序列向二维结构的转换,保留时间-尺度关联信息;然后,引入融合深度可分卷积的金字塔注意力机制(Pyramid Convolutional Block Attention Module integrated with Depthwise Separable Convolution,P-CBAM-DSC),提升模型对故障区域的聚焦能力与上下文表达能力;最终,通过全局特征聚合分类器(Global Feature Aggregation Classifier,GFA-C)实现高效的端到端故障识别。在DCASE2023公开声音数据集与自建滚动轴承声纹平台上的实验结果表明,所提方法在准确率、模型轻量化与推理效率方面均优于主流轻量模型,展现出良好的诊断性能、噪声鲁棒性与边缘部署适应性。
To address the strong nonstationarity and significant noise interference commonly present in acoustic signals of rotating machinery,as well as the limitations of existing methods,including insufficient capability in temporal-scale feature modeling,reliance on handcrafted time-frequency transformations,and excessive model complexity that hinders edge deployment,a lightweight end-to-end fault diagnosis model was proposed that integrates Multiscale Dynamic Attention and 1D-2D convolutional Fusion Network(MDAF-Net).The proposed model comprised four key components.First,a Multiscale Dynamic Weighting Feature Extractor(MDW-FE)was constructed to enhance the perception of nonstationary acoustic patterns through the combination of multiscale convolutional kernels and adaptive weighting mechanisms.Second,a Reshaped Multiscale Projection(RMP)layer was designed to transform one-dimensional sequences into two-dimensional structures,thereby preserving temporal-scale dependencies.Third,a Pyramid Convolutional Block Attention Module integrated with Depthwise Separable Convolution(P-CBAM-DSC)was introduced to improve the model′s capability in focusing on fault-sensitive regions and capturing contextual semantics.Finally,a Global Feature Aggregation Classifier(GFA-C)enables efficient and accurate end-to-end fault identification.Experimental results on the public DCASE2023 dataset and a self-built rolling bearing acoustic benchmark demonstrate that the proposed method outperforms mainstream lightweight models in terms of diagnostic accuracy,model compactness,and inference efficiency,while exhibiting excellent performance in noise robustness and suitability for edge deployment.
作者
何新荣
杜小泽
谭锐
蒋国安
徐超
HE Xinrong;DU Xiaoze;TAN Rui;JIANG Guoan;XU Chao(China Energy Nanjing Electric Power Test&Research Co.,Ltd.,Nanjing 210023,China;North China Electric Power University,Beijing 102206,China)
出处
《现代制造工程》
北大核心
2026年第2期143-150,共8页
Modern Manufacturing Engineering
基金
国家能源集团科技项目(GJNY-23-68)
国家能源集团科学技术研究院有限公司科技项目(DY2025Y01)。
关键词
旋转机械
故障诊断
声学信号
轻量化网络
1D-2D卷积建模
多尺度动态注意力
rotating machinery
fault diagnosis
acoustic signal
lightweight network
1D-2D convolutional modeling
multiscale dynamic attention
