Scholarlink

<p>To address the low detection accuracy (&lt; 70%) of existing methods for complex faults (d03, d09, d15) in the Tennessee Eastman process, this paper proposes an evolutionary weighted feature influence factor (WFIF) method. The proposed method addresses the limitation that most existing approaches mainly focus on discriminative power and dimensionality, while paying insufficient attention to the stability of feature subsets in repeated selection. Stability is defined as the algorithm’s ability to produce highly consistent subsets across different data partitions. Comparative experiments with six existing feature selection methods show that the evolutionary WFIF method improves the F1-score for the three complex faults from below 73% to over 85%, while raising the average F1-score for all 21 TE process faults to 95.79%. Regarding dimensionality, it reduces the average number of selected features to 9.5% of the original 52 variables. In terms of stability (quantification of the dimension size of feature subset by standard deviation), the standard deviation of the selected dimension size is zero for 20 out of 21 faults, with only fault d12 having a deviation of 1, which is significantly lower than that of other methods. Cross-classifier validation further demonstrates the excellent model generalization capability of the selected feature subsets.</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

An evolutionary weighted feature influence factor feature selection method for fault detection in the Tennessee Eastman complex chemical process

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月27日

查看全文
  • Dongliang Li,
  • Yulong Xue,
  • Jinzhou Fu,
  • Tian Tan,
  • Yi Yang,
  • Juchen Nie

摘要

To address the low detection accuracy (< 70%) of existing methods for complex faults (d03, d09, d15) in the Tennessee Eastman process, this paper proposes an evolutionary weighted feature influence factor (WFIF) method. The proposed method addresses the limitation that most existing approaches mainly focus on discriminative power and dimensionality, while paying insufficient attention to the stability of feature subsets in repeated selection. Stability is defined as the algorithm’s ability to produce highly consistent subsets across different data partitions. Comparative experiments with six existing feature selection methods show that the evolutionary WFIF method improves the F1-score for the three complex faults from below 73% to over 85%, while raising the average F1-score for all 21 TE process faults to 95.79%. Regarding dimensionality, it reduces the average number of selected features to 9.5% of the original 52 variables. In terms of stability (quantification of the dimension size of feature subset by standard deviation), the standard deviation of the selected dimension size is zero for 20 out of 21 faults, with only fault d12 having a deviation of 1, which is significantly lower than that of other methods. Cross-classifier validation further demonstrates the excellent model generalization capability of the selected feature subsets.