Intelligent hardware acceleration optimization method for real-time junction temperature evaluation of power IGBT modules

Power Insulated Gate Bipolar Transistor modules are crucial in applications such as electric vehicles and renewable energy systems, where junction temperature significantly impacts reliability and lifespan. While data-driven multi-parameter methods enhance junction temperature prediction accuracy and robustness, the computational complexity of AI models poses challenges for deployment in power systems, particularly on embedded platforms. This study proposes a system integrating algorithm optimization and hard- ware acceleration to achieve real-time junction temperature prediction for IGBT modules. Utilizing the saturation collector-emitter voltage, collector current, and case temperature as inputs, the system incorporates the proposed approximate prior vector inner product method for optimizing vector operations in a BP neural network and employs a multi-stage neural processing unit on the Zynq7000 platform. Experimental results demonstrate that, during the online junction temperature inference stage, the proposed system achieves an average junction temperature prediction error of 2 K, with 1000 inferences completed in 0.917 ms, achieving a 34.85-fold energy efficiency improvement over comparable platforms. The neural network training is performed offline and is not included in the reported time and energy consumption. This work provides an efficient solution for embedding artificial intelligence in power electronics systems from both algorithm and hardware optimization perspectives, advancing reliability and lifespan management while supporting edge computing deployment in energy and electrical engineering fields.

Microelectronics Reliability

Microelectronics Reliability

179

1872-941X

N/A

N/A

14