Prediction Method of Driving Strategy of High-Power IGBT Module Based on MEA-BP Neural Network

Zhaoliang Meng, Yuan Yang, Yong Gao, Shengsheng Ai, Yue Zhang, Yaru Lv, Zetao Zhang, Yang Wen, Lei Wu, Ping Zhang, John Douglas ThomsonIEEE Access 2020 · DOI

Introduction

In high-power electronics applications such as electric locomotives, smart grids, and industrial drives, the Insulated Gate Bipolar Transistor (IGBT) serves as a critical switching component. However, driving these modules effectively requires managing a difficult trade-off: switching speed versus device stress. Rapid switching reduces energy loss but induces significant current and voltage overshoots that can damage the module. Conversely, slow switching protects the device but leads to excessive heat generation through switching losses. Traditional gate drivers use fixed resistors, which represent a static compromise that is rarely optimal across all operating conditions.

The Challenge of Active Gate Driving

Active Gate Drivers (AGD) offer a solution by allowing variable gate resistance during the switching process. By adjusting resistance dynamically, it is possible to "soften" the transitions where spikes occur while maintaining high speed elsewhere. However, determining the optimal resistance values for every microsecond of a non-linear switching event is an immense optimization challenge.

Basic switching current paths for turn-on (a) and turn-off (b) in an active gate driver circuit.

As shown in the basic driver paths above, the current must be carefully managed through the gate resistors ($R_{gon}$ and $R_{goff}$). For high-power modules, manual tuning of these parameters is tedious and often fails to find the global optimum.

Proposed Method: MEA-BP Neural Network Optimization

Researchers have proposed a data-driven approach using a three-stage variable gate resistance strategy. By employing different resistance values ($R_{g1}, R_{g2}, R_{g3}$) for distinct phases of the switching transition, the driver can be fine-tuned for maximum efficiency and safety.

To predict these optimal values, the study utilizes a Backpropagation (BP) Neural Network enhanced by the Mind Evolutionary Algorithm (MEA). While standard BP networks are prone to getting stuck in local sub-optimal solutions, the MEA mimics human social learning—using "similartaxis" (local search) and "dissimilation" (global search)—to ensure the model identifies the absolute best driving strategy.

The multi-stage variable gate resistance driver architecture controlled by the optimized MEA-BP strategy.

The hardware implementation involves a controller managing a bank of MOSFETs ($Q_1$ through $Q_6$) and corresponding resistors ($R_{g1}$ through $R_{g6}$), as illustrated in the architecture above. This allows the system to switch between different resistance paths in real-time based on the neural network's predictions.

Experimental Results and Insights

The model was trained and validated using a high-power experimental test bench featuring 4500V/900A Mitsubishi Planar-IGBT modules. The results demonstrated that the MEA-BP optimized strategy significantly outperformed conventional fixed-gate drivers (CGD):

  • Overshoot Current ($\Delta I_c$): Reduced by 59.31%.
  • Turn-on Energy Loss ($E_{on}$): Reduced by 36.99%.
  • Turn-on Time ($t_{on}$): Reduced by 46.38%.
  • Overall Performance: The Comprehensive Evaluation Method (CEM) score improved by 65.65%.

Interestingly, the study found that for Planar-IGBTs, the turn-off process is largely insensitive to gate resistance changes. This allowed the researchers to simplify the driver by focusing optimization primarily on the turn-on phase. Furthermore, the optimized strategies remained robust across a wide temperature range (25°C to 125°C) and did not increase electromagnetic interference (EMI).

Conclusion

By combining evolutionary algorithms with neural networks, this research provides a robust framework for automating the design of high-performance gate drivers. The massive reductions in both stress (overshoot) and waste (energy loss) pave the way for more efficient and reliable power electronics. Future work will look toward "online" adaptive learning and applying these techniques to next-generation Wide-Bandgap (WBG) semiconductors like Silicon Carbide (SiC) MOSFETs.

Summary updated

BibTeX
@article{meng2020prediction,
  author = {Zhaoliang Meng and Yuan Yang and Yong Gao and Shengsheng Ai and Yue Zhang and Yaru Lv and Zetao Zhang and Yang Wen and Lei Wu and Ping Zhang and John Douglas Thomson},
  title = {{Prediction Method of Driving Strategy of High-Power IGBT Module Based on MEA-BP Neural Network}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2995601},
}
Figure