R-PODID aims to develop the technologies and methodologies required to improve the reliability in power electronics circuits and systems based on III/V-semiconductor devices like gallium-nitride (GaN) and silicon-carbine (SiC). GaN and SiC devices will be the standard devices for power electronic systems of the future thanks to their electric and thermal performance, but they suffer reliability limitations. Moreover, periodic shutdowns of industrial plants for maintenance purposes set a considerable industrial cost, while a modern AI-based prediction of the failure may lower the exercise costs and trigger novel business models. R-PODID targets all the above scenarios by developing models, circuits, and systems for improving the reliability of power electronics systems.
The project's basic idea is to incorporate and combine artificial intelligence (AI) algorithms with state-of-the-art physical models to realize compact failure models that can be deployed on resource-constrained embedded systems commonly used in power applications, allowing for short-term fault prediction and real-time residual lifetime estimation. To achieve the final goal, the project has to face interdisciplinary challenges and set ambitious objectives.
Failure analysis is typically characterized by a reduced training dataset, therefore novel methodologies for developing AI models from sparse datasets must be conceived. Advanced AI learning techniques and combination with compressive sensing, will be investigated for this purpose. AI models are usually run in the cloud, on very powerful servers. However, the applications foreseen in this project require the local deployment of the failure model on embedded devices, which are usually resource-constrained. This is especially true in the automotive context, which is targeted by two Use Cases in R-PODID. The development of efficient methodologies for the deployment of the AI, or the AI-enhanced, model into the embedded device is another objective of the R-PODID project. The project will target the most used embedded architectures: a RISC-V-based system with safety capabilities for industrial applications and an ARM-cortex MCU with ASIL-D certifications for automotive applications. Finally, the AI-based models require a constant update of information about the current behavior of the monitored system, therefore specific and performing sensors must be developed to gather a great amount of data with the highest possible level of accuracy, given the critical nature of the applications. Moreover, in the case of federated learning over large industrial plants, secure and reliable communication systems must be defined. The last challenge is the integration of all the above-described methodologies and technologies into power-converter architectures without impinging the converter's original performance.
The achievement of the R-PODID goal will allow for an automated, cloudless, short-term fault prediction for electric drives, power modules, and power devices, that can be integrated into power converters. Thereby, electrical and mechanical faults of machines and of the power converters driving them will become predictable within a limited prediction horizon of 12-24 hours. This will enable a power-saving shutdown of a larger number of production machines during idle times, because a looming failure during the next power-on cycle can be reliably foreseen. It will also enable reliable mitigation of dangerous faults in critical applications like automotive.