IGBT Module DPT Efficiency Enhancement via Multimodal Fusion Networks and Graph Convolution Networks

TY - JOUR

T1 - IGBT Module DPT Efficiency Enhancement via Multimodal Fusion Networks and Graph Convolution Networks

AU - Zhang, Xiaotian

AU - Hu, Yihua

AU - Zhang, Jingwei

AU - Esfahani, Mohammad Nasr

AU - Tilford, Tim

AU - Stoyanov, Stoyan

N1 - This is an author-produced version of the published paper. Uploaded in accordance with the University’s Research Publications and Open Access policy.

PY - 2024/3/7

Y1 - 2024/3/7

N2 - The dynamic electrical characteristics of insulated-gate bipolar transistor (IGBT) are of great significance in practical high-power electrical applications and are usually evaluated through double pulse test (DPT). However, DPTs of IGBTs under various working conditions are time-consuming and laborious. Traditional estimation methods are based on detailed physical parameters and complex formula calculations, making deployment process challenging. This article proposes a novel DPT efficiency enhancement method based on graph convolution network (GCN) and feature fusion technology, which can estimate and supplement switching transient waveforms of all working conditions. Thereby, dynamic electrical characteristics of the IGBT are obtained by estimated waveforms of DPT. This method proposes a multimodal attention fusion network to capture and fuse the features of switching transient waveforms between different positions thereby improving the expressive power and performance of the model. Moreover, this method is novel in that it is the first to utilize GCN to embed DPT data under multiple working conditions into a graph structure, which can use the graph structure information to fuse the features of spatially correlated working conditions data to obtain reliable estimation results. The method has been verified to be effective and accurate on real dataset collected on two batches of IGBTs.

AB - The dynamic electrical characteristics of insulated-gate bipolar transistor (IGBT) are of great significance in practical high-power electrical applications and are usually evaluated through double pulse test (DPT). However, DPTs of IGBTs under various working conditions are time-consuming and laborious. Traditional estimation methods are based on detailed physical parameters and complex formula calculations, making deployment process challenging. This article proposes a novel DPT efficiency enhancement method based on graph convolution network (GCN) and feature fusion technology, which can estimate and supplement switching transient waveforms of all working conditions. Thereby, dynamic electrical characteristics of the IGBT are obtained by estimated waveforms of DPT. This method proposes a multimodal attention fusion network to capture and fuse the features of switching transient waveforms between different positions thereby improving the expressive power and performance of the model. Moreover, this method is novel in that it is the first to utilize GCN to embed DPT data under multiple working conditions into a graph structure, which can use the graph structure information to fuse the features of spatially correlated working conditions data to obtain reliable estimation results. The method has been verified to be effective and accurate on real dataset collected on two batches of IGBTs.

KW - Behavioral sciences

KW - Double pulse test (DPT)

KW - Employee welfare

KW - Estimation

KW - feature fusion

KW - graph convolutional network (GCN)

KW - Insulated gate bipolar transistors

KW - insulated-gate bipolar transistor (IGBT)

KW - Integrated circuit modeling

KW - Switches

KW - Transient analysis

UR - http://www.scopus.com/inward/record.url?scp=85187369841&partnerID=8YFLogxK

U2 - 10.1109/TIE.2024.3368165

DO - 10.1109/TIE.2024.3368165

M3 - Article

AN - SCOPUS:85187369841

SN - 0278-0046

SP - 1

EP - 12

JO - IEEE Transactions on Industrial Electronics

JF - IEEE Transactions on Industrial Electronics

ER -