报告详情
  1. 活动首页
  2. 报告清单
  3. 报告详情
LDPMNet: A Deployment-Oriented Multi-Bin Prototypical Network with Depthwise Separable and Partial Convolutions for Lightweight Cross-Component Few-Shot Fault Diagnosis
编号：75 访问权限：仅限参会人 更新：2026-09-23 10:20:50 浏览：1次 口头报告

报告开始：暂无开始时间（Asia/Shanghai）

报告时间：暂无持续时间

所在会场：[暂无会议] [暂无会议段]

暂无文件

摘要
Cross-component few-shot mechanical fault diagnosis suffers from limited labeled target samples, significant spectral discrepancies among components, and high deployment costs on resource-constrained SoCs. Existing studies primarily optimize diagnostic accuracy, but many still rely on computationally intensive feature extractors. Moreover, model complexity, hardware resource utilization, and latency are not consistently quantified for specified SoC targets. To address these issues, this paper proposes a deployment-oriented Lightweight Depthwise-Separable and Partial-Convolution Multi-Bin Prototypical Network (LDPMNet). Its encoder employs depthwise separable convolution (DSC) to decouple feature-axis filtering from channel mixing, thereby reducing parameter and computational costs. One-dimensional partial convolution (PConv) processes only a subset of channels while retaining the remaining features, further reducing redundant computation and preserving feature information. Multi-bin adaptive average pooling preserves ordered regional statistics along the encoded feature axis to produce fixed-dimensional embeddings. Finally, class prototypes are constructed from support embeddings, and query samples are classified according to their squared Euclidean distances to these prototypes. Experiments are conducted on three cross-component transfer directions under unified 5-way 1/3/5-shot settings. LDPMNet achieves a macro-averaged accuracy of 92.06% across the nine task-shot configurations. FP32 encoder HLS synthesis estimated BRAM18K and LUT utilization at 37.50% and 6.99% of the XC7Z020 capacities, respectively, providing preliminary resource estimates for subsequent system integration.
 
关键词
Cross-component few-shot, lightweight network, DSC, PConv, multi-bin pooling
报告人
Zhenkai Meng
Graduate Student Anhui University

稿件作者
Zhenkai Meng Anhui University
Juan Xu Anhui University
永斌 刘 安徽大学
俊博 赵 安徽大学
发表评论
验证码 看不清楚，更换一张
全部评论
重要日期

  • 会议日期

    11月06日

    2026

    至

    11月08日

    2026

  • 10月15日 2026

    初稿截稿日期

主办单位
IEEE Instrumentation and Measurement Society
承办单位
Sichuan University
联系方式
登录查看完整联系方式
历届会议
移动端
在手机上打开
小程序
打开微信小程序
客服
扫码或点此咨询