Scholarlink

<p>To address the nonlinear contact problem during the hole-finding phase of robotic peg-in-hole assembly on flexible support platforms, which significantly&#xa0;affects&#xa0;hole-centering accuracy, this study proposes a deviation prediction method based on time-series force perception, optimized through the Chinese Pangolin Optimizer–Convolutional Neural Network-Bidirectional Long Short-Term Memory (CPO-CNN-BiLSTM). This method utilizes CNN and BiLSTM to extract contact force features and capture bidirectional temporal dependencies, while the CPO is introduced to optimize key network parameters, thereby constructing a prediction model for relative hole center deviation. Comparative experiments were conducted on two flexible support platforms with different stiffness coefficients. The results demonstrate that the proposed method outperforms four reference architecture models (CNN, LSTM, BiLSTM, and CNN-BiLSTM) in terms of performance metrics, exhibiting superior generalization capability and stability. Furthermore, in hole-finding experiments under both stiffness conditions, a success rate of 91.7% was achieved, validating the method’s applicability within a certain flexibility range. The findings provide a feasible technical solution for force perception and deviation compensation in flexible assembly environments. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning-based hole-finding for robotic peg-in-hole assembly with time-series force data on flexible support platform

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Jiaming Xiong,
  • Haibin Huang,
  • Jiahao He,
  • Zihao Ye

摘要

To address the nonlinear contact problem during the hole-finding phase of robotic peg-in-hole assembly on flexible support platforms, which significantly affects hole-centering accuracy, this study proposes a deviation prediction method based on time-series force perception, optimized through the Chinese Pangolin Optimizer–Convolutional Neural Network-Bidirectional Long Short-Term Memory (CPO-CNN-BiLSTM). This method utilizes CNN and BiLSTM to extract contact force features and capture bidirectional temporal dependencies, while the CPO is introduced to optimize key network parameters, thereby constructing a prediction model for relative hole center deviation. Comparative experiments were conducted on two flexible support platforms with different stiffness coefficients. The results demonstrate that the proposed method outperforms four reference architecture models (CNN, LSTM, BiLSTM, and CNN-BiLSTM) in terms of performance metrics, exhibiting superior generalization capability and stability. Furthermore, in hole-finding experiments under both stiffness conditions, a success rate of 91.7% was achieved, validating the method’s applicability within a certain flexibility range. The findings provide a feasible technical solution for force perception and deviation compensation in flexible assembly environments.