Scholarlink

<p>Friction is one of the primary factors affecting the tracking performance of a machine tool feed system. Establishing an accurate friction model is the key to achieving friction compensation. Although neural networks trained on friction data can make accurate predictions, they cannot adapt to friction characteristic changes that occur over time because their model parameters are fixed. In this work, a sample acquisition scheme with changing friction characteristics is first designed. Then, an adaptive modeling method using dual neural networks is proposed. The method constructs two networks with the same structure, Net-P and Net-T, which carry out the prediction and training tasks, respectively. While Net-P performs prediction in real-time, Net-T can continuously update its parameters with the latest friction data. The parameters of Net-T are then copied to Net-P to adapt to the friction characteristic changes. In addition, a classifier-based sample pool is proposed to enable the model to learn the latest friction characteristics, while preventing overfitting. The experimental results show that the proposed method can successfully adapt to variation in the friction characteristics, with a root mean square error (RMSE) of 0.034 Nm and a maximum absolute error (MAE) of 0.248 Nm. Compared with those of the three comparison models, the RMSEs of the proposed model are 13.7%, 70.8%, and 73.9% smaller, and the MAEs are 40.1%, 45.2%, and 47.2% smaller. Moreover, the prediction results of the proposed model are smoother than those of the compared adaptive models and less likely to cause system vibration.</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

Adaptive friction modeling in feeding systems based on dual neural networks

  • 期刊论文

  • 发表日期：

    2025年3月11日

查看全文
  • Dehai Huang,
  • Jianzhong Yang,
  • Huicheng Zhou,
  • Guangda Xu

摘要

Friction is one of the primary factors affecting the tracking performance of a machine tool feed system. Establishing an accurate friction model is the key to achieving friction compensation. Although neural networks trained on friction data can make accurate predictions, they cannot adapt to friction characteristic changes that occur over time because their model parameters are fixed. In this work, a sample acquisition scheme with changing friction characteristics is first designed. Then, an adaptive modeling method using dual neural networks is proposed. The method constructs two networks with the same structure, Net-P and Net-T, which carry out the prediction and training tasks, respectively. While Net-P performs prediction in real-time, Net-T can continuously update its parameters with the latest friction data. The parameters of Net-T are then copied to Net-P to adapt to the friction characteristic changes. In addition, a classifier-based sample pool is proposed to enable the model to learn the latest friction characteristics, while preventing overfitting. The experimental results show that the proposed method can successfully adapt to variation in the friction characteristics, with a root mean square error (RMSE) of 0.034 Nm and a maximum absolute error (MAE) of 0.248 Nm. Compared with those of the three comparison models, the RMSEs of the proposed model are 13.7%, 70.8%, and 73.9% smaller, and the MAEs are 40.1%, 45.2%, and 47.2% smaller. Moreover, the prediction results of the proposed model are smoother than those of the compared adaptive models and less likely to cause system vibration.