Scholarlink

System identification is a crucial component in industrial control systems, with its core objective being the construction of accurate dynamic models based on input-output data. For first-order plus dead time (FOPDT) models, traditional identification methods often struggle with limited robustness and adaptability under complex operating conditions and strong disturbances. To address these challenges, the paper proposes a novel parameter identification framework, TCN-MTL-FOPDT, which integrates Temporal Convolutional Networks (TCN) and Multi-Task Learning (MTL) to enhance estimation accuracy and generalization. The framework employs Model Predictive Control (MPC) to dynamically generate input-output data, utilizes TCNs to extract essential temporal features, and adopts a multi-task learning strategy to jointly optimize the main task (FOPDT parameter estimation) and an auxiliary task (future output prediction) via shared feature representations. Furthermore, a dynamic weighting mechanism is introduced to balance the training between tasks. Experimental results demonstrate that the proposed TCN-MTL-FOPDT framework outperforms traditional approaches such as ARX-LS and LSSVM, achieving an average improvement of approximately 18.6% in parameter estimation accuracy and demonstrating superior robustness across varying noise levels.

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

System Identification Based on Temporal Convolutional Networks and Multi-task Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jian Chen,
  • Haiwei Pan,
  • Zhenzhong Xu,
  • Yong Pan

摘要

System identification is a crucial component in industrial control systems, with its core objective being the construction of accurate dynamic models based on input-output data. For first-order plus dead time (FOPDT) models, traditional identification methods often struggle with limited robustness and adaptability under complex operating conditions and strong disturbances. To address these challenges, the paper proposes a novel parameter identification framework, TCN-MTL-FOPDT, which integrates Temporal Convolutional Networks (TCN) and Multi-Task Learning (MTL) to enhance estimation accuracy and generalization. The framework employs Model Predictive Control (MPC) to dynamically generate input-output data, utilizes TCNs to extract essential temporal features, and adopts a multi-task learning strategy to jointly optimize the main task (FOPDT parameter estimation) and an auxiliary task (future output prediction) via shared feature representations. Furthermore, a dynamic weighting mechanism is introduced to balance the training between tasks. Experimental results demonstrate that the proposed TCN-MTL-FOPDT framework outperforms traditional approaches such as ARX-LS and LSSVM, achieving an average improvement of approximately 18.6% in parameter estimation accuracy and demonstrating superior robustness across varying noise levels.