Scholarlink

<p>This study presents a surrogate-assisted, many-objective optimization framework for the seals and bearings of a seven-stage BB5 centrifugal pump. The optimization aimed to reduce bearing power loss, total seal leakage, amplification factor, and the ratio of the vibration amplitude to the component clearances, while increasing the stability metrics such as logarithmic decrement and separation margin. Seal leakage and the rotordynamic coefficients were obtained using the bulk-flow model, whereas the bearing performance was evaluated using the Sommerfeld number-based nondimensional data. The neural-network surrogate models were introduced to eliminate the computational bottleneck of repeatedly solving seal equations within the optimization loop. A multi-layer perceptron (MLP) predicts leakage at the operating speed, and a multi-head Deep Operator Network (DeepONet) predicts speed-dependent dynamic coefficients. Embedding these surrogates into the evaluation loop accelerates a single generation by approximately 80.2-fold compared to the baseline numerical workflow, reducing the average per-generation cost from 405 seconds to 5.05 seconds on the tested hardware. Optimization was carried out using the Unified Non-dominated Sorting Genetic Algorithm&#xa0;III (U-NSGA-III). Compared to the baseline configuration, the resulting Pareto-optimal designs reduced the total leakage by more than 70% together with lower bearing power loss, while satisfying the rotordynamic stability requirements, including the minimum logarithmic decrement, amplification factor limits, minimum component clearances, and separation margins. These results show that the surrogate-assisted optimization method can identify the non-intuitive combinations of seal tapers, inlet swirl ratios, and bearing types that jointly enhance the efficiency and dynamic robustness in multistage pumps.</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

Multi-Objective Optimization of Seals and Bearings in Multistage Pumps via Neural Network Surrogate Models

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Jeongin Lee,
  • Donghyun Lee,
  • Jin Woo Choi,
  • Junho Suh

摘要

This study presents a surrogate-assisted, many-objective optimization framework for the seals and bearings of a seven-stage BB5 centrifugal pump. The optimization aimed to reduce bearing power loss, total seal leakage, amplification factor, and the ratio of the vibration amplitude to the component clearances, while increasing the stability metrics such as logarithmic decrement and separation margin. Seal leakage and the rotordynamic coefficients were obtained using the bulk-flow model, whereas the bearing performance was evaluated using the Sommerfeld number-based nondimensional data. The neural-network surrogate models were introduced to eliminate the computational bottleneck of repeatedly solving seal equations within the optimization loop. A multi-layer perceptron (MLP) predicts leakage at the operating speed, and a multi-head Deep Operator Network (DeepONet) predicts speed-dependent dynamic coefficients. Embedding these surrogates into the evaluation loop accelerates a single generation by approximately 80.2-fold compared to the baseline numerical workflow, reducing the average per-generation cost from 405 seconds to 5.05 seconds on the tested hardware. Optimization was carried out using the Unified Non-dominated Sorting Genetic Algorithm III (U-NSGA-III). Compared to the baseline configuration, the resulting Pareto-optimal designs reduced the total leakage by more than 70% together with lower bearing power loss, while satisfying the rotordynamic stability requirements, including the minimum logarithmic decrement, amplification factor limits, minimum component clearances, and separation margins. These results show that the surrogate-assisted optimization method can identify the non-intuitive combinations of seal tapers, inlet swirl ratios, and bearing types that jointly enhance the efficiency and dynamic robustness in multistage pumps.