Scholarlink

<p>Power electronic converters (PECs) are crucial to modern electrical systems, facilitating efficient energy conversion in applications from renewable energy systems, electric vehicles, to industrial drives. To mitigate risks of extended downtime, financial losses, and health hazards associated to failure of critical PEC components like Insulated Gate Bipolar Transistors (IGBTs) while in operation, several data-driven machine learning models have been proposed in literature to estimate their remaining useful life (RUL) in order to improve system reliability and perform predictive maintenance of these critical components of the power electronic converter based on National Aeronautics and Space Administration (NASA’s) accelerated aging datasets. However, reliance on recurrence in machine learning models in literature mainly lack robustness and accuracy due to the size and patterns of the NASA accelerated aging dataset. This paper develops a novel data-driven approach to this issue. For the IGBTs, a new Transformer model, the Encoder-Transformer, is developed for RUL multivariate time-series forecasting. It employs the multi-head attention mechanism to capture long-term dependencies and outperforms traditional RNN architectures—Simple Recurrent Neural Network (Simple RNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM). The Encoder-Transformer was found to be superior compared to Simple RNN, GRU, and LSTM with RMSE of 53.81, 156.14, 166.32 and 139.89 for IGBT #3 generalization; 14.90, 105.14, 78.84 and 92.78 for IGBT #4; and 37.76, 62.34, 111.69 and 114.92 for IGBT #5 respectively based on RUL with units in seconds, trained on NASA’s IGBT #2 dataset using the steady-state data precursors. The proposed data-driven method produces robust, scalable solutions for component level prognostics.</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

Development of encoder-multivariate time-series transformer for remaining useful life prediction of IGBTs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月6日

查看全文
  • Kofi Addo Annan,
  • Francis Boafo Effah,
  • Emmanuel Asuming Frimpong

摘要

Power electronic converters (PECs) are crucial to modern electrical systems, facilitating efficient energy conversion in applications from renewable energy systems, electric vehicles, to industrial drives. To mitigate risks of extended downtime, financial losses, and health hazards associated to failure of critical PEC components like Insulated Gate Bipolar Transistors (IGBTs) while in operation, several data-driven machine learning models have been proposed in literature to estimate their remaining useful life (RUL) in order to improve system reliability and perform predictive maintenance of these critical components of the power electronic converter based on National Aeronautics and Space Administration (NASA’s) accelerated aging datasets. However, reliance on recurrence in machine learning models in literature mainly lack robustness and accuracy due to the size and patterns of the NASA accelerated aging dataset. This paper develops a novel data-driven approach to this issue. For the IGBTs, a new Transformer model, the Encoder-Transformer, is developed for RUL multivariate time-series forecasting. It employs the multi-head attention mechanism to capture long-term dependencies and outperforms traditional RNN architectures—Simple Recurrent Neural Network (Simple RNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM). The Encoder-Transformer was found to be superior compared to Simple RNN, GRU, and LSTM with RMSE of 53.81, 156.14, 166.32 and 139.89 for IGBT #3 generalization; 14.90, 105.14, 78.84 and 92.78 for IGBT #4; and 37.76, 62.34, 111.69 and 114.92 for IGBT #5 respectively based on RUL with units in seconds, trained on NASA’s IGBT #2 dataset using the steady-state data precursors. The proposed data-driven method produces robust, scalable solutions for component level prognostics.