Scholarlink

<p>An improved particle swarm optimization (IPSO) method is proposed to train a back-propagation (BP) neural network for ABS fault classification. IPSO integrates three operators: (i) a dynamic inertia-weight schedule to balance global exploration and local exploitation; (ii) time-varying cognitive and social learning factors to adapt self-/swarm-learning strengths across iterations; and (iii) a variance-dispersion mutation to escape local minima and mitigate premature convergence. These operators jointly optimize the BP network’s weights and biases. An ABS fault co-simulation on CarSim/Simulink yields 720 labeled samples (nine classes), and an in-vehicle dataset contains 2,160 labeled samples. Data are split with class-wise stratified 6:2:2 partitioning under a fixed random seed, and features are min–max normalized with statistics computed on the training set only to prevent leakage. On simulation data, IPSO-BP improves overall accuracy by 4.1% over PSO-BP. On in-vehicle data, IPSO-BP outperforms GA-BP, IPSO-SVM, and GA-SVM by 4.9%, 6.5%, and 9.5%, respectively, demonstrating superior accuracy. These results indicate that the proposed IPSO operators substantially enhance PSO’s search dynamics and yield a more accurate and reliable BP-based diagnostic model for ABS.</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

Automotive ABS Fault Classification and Identification Based on IPSO-BP Neural Network

  • 期刊论文

  • 发表日期：

    2026年1月23日

查看全文
  • Tao Zhang,
  • Zhifeng Zhu,
  • Guotai Ji,
  • Cheng Qian,
  • Ke Yang,
  • Bohua Cai,
  • Yong Yao

摘要

An improved particle swarm optimization (IPSO) method is proposed to train a back-propagation (BP) neural network for ABS fault classification. IPSO integrates three operators: (i) a dynamic inertia-weight schedule to balance global exploration and local exploitation; (ii) time-varying cognitive and social learning factors to adapt self-/swarm-learning strengths across iterations; and (iii) a variance-dispersion mutation to escape local minima and mitigate premature convergence. These operators jointly optimize the BP network’s weights and biases. An ABS fault co-simulation on CarSim/Simulink yields 720 labeled samples (nine classes), and an in-vehicle dataset contains 2,160 labeled samples. Data are split with class-wise stratified 6:2:2 partitioning under a fixed random seed, and features are min–max normalized with statistics computed on the training set only to prevent leakage. On simulation data, IPSO-BP improves overall accuracy by 4.1% over PSO-BP. On in-vehicle data, IPSO-BP outperforms GA-BP, IPSO-SVM, and GA-SVM by 4.9%, 6.5%, and 9.5%, respectively, demonstrating superior accuracy. These results indicate that the proposed IPSO operators substantially enhance PSO’s search dynamics and yield a more accurate and reliable BP-based diagnostic model for ABS.