0016 Dynamic Electromagnetic-Electro-Thermal Coupled Modeling of Power Conversion System During Load Cycles
Hsien.-Chie Cheng, Yan-Chen Liu and Cong-Jun Huang
The power electronics market in the automotive industry and energy industry witnesses an explosive growth due to growing need of eco-friendly vehicles and drastic climate change and global warming. Power electronics inside a power conversion system, such as insulated-gate bipolar transistor (IGBT) and metal-oxide semiconductor field effect transistor (MOSFET), are the key device determining the efficiency of power conversion. Today, the rapid development of power electronics industry closely follows the trend of the increase in power rating, switching frequency and the decrease of the size, which would unavoidably give rise to high power dissipation density. The high power dissipation density together with harsh operating environment would result in high device temperature, potentially leading to electrical degradation and even breakdown, and thermal-mechanical failure. The objective of the study is to introduce an efficient and effective dynamic multi-physics modeling framework to predict the Electromagnetic-Electro-Thermal coupled behavior of power conversion systems like inverters and converters during load cycles. This modeling framework combines an integrated electromagnetic circuit (E-C) model for exploration of the parasitics parameters and their influence on the switching transients and power losses, and a modified dynamic fully coupled electro-thermal (E-T) model for estimation of instantaneous power loss dependence on instantaneous temperature and device junction temperature. The modified E-T model incorporates a compact thermal model based on Foster network, constructed via parametric computational fluid dynamics (CFD) thermal analysis, and a simple power (P)-temperature (T) relationship, in replace of the time-consuming circuit simulation often used in the conventional dynamic electro-thermal model. The proposed modeling framework is tested on a three-phase inverter operating with a 180-degree conduction mode using six power MOSFET devices as switching devices for brushless DC motor drive during six-step commutation. To validate the proposed E-C model and CFD thermal model, double pulse test and IR thermography experiments are performed, respectively. Besides, the effectiveness of the proposed modeling framework is demonstrated by comparing with the calculated transient device junction temperature with that of the conventional dynamic E-T model. At last, parametric analysis is performed to examine the effects of several key factors on the thermal performance of the inverter.
| |
0012 車用電子元件之脫層失效分析
Mei-Ling Wu and Jia-Shen Lan
Please refer to the attached file.
| |
0010 高功率模組封裝製程與可靠度試驗耦合效應之模擬分析與驗證
昌駿 李, 繼元 許 and 元呈 黃
由於油電混合/純電能驅動個人行動載具與大眾軌道運輸系統發展快速,而高功率模組作為驅動系統之關鍵零組件,因此高功率模組製程翹曲量與熱循環可靠度評估與測試是必須的。有鑒於此,本論文提出一功率模組製程兼熱循環之耦合力學效應有限元素方法,並考慮SAC305焊錫黏塑性行為與環氧樹脂之化學收縮特性影響。此模擬方法考慮功率模組製程所引致翹曲與殘留應力,更實際探討後續溫度與功率循環可靠度負載作用下之失效位置與可靠度估算的方法。模擬顯示功率模組AIN陶瓷層亦或是焊錫層為主要失效之可能性。因此,依據實際可行性對功率模組製程翹曲量與熱循環可靠度予以優化,針對環氧樹脂材料機械性質與陶瓷材料種類與厚度,進行中央合成設計與參數化分析。研究結果指出,環氧樹脂材料性質之楊氏係數與熱膨脹係數優化可減緩製程時熱膨脹不匹配,致使功率模組翹曲量與熱循環可靠度具有全面性之改善。而陶瓷材料選擇與厚度參數化分析顯示,具有高熱膨脹係數與厚度薄化之陶瓷可大幅地提高焊錫層之可靠度。
| |
0052 Deep Learning of the SSL Luminaire Spectral Power Distribution under Multiple Degradation Mechanisms by Hybrid kNN algorithm
Cadmus Yuan and G.Q. Zhang
Solid-state lighting (SSL) is a technology evolution for lighting applications. The high brightness, small size, and white light LED light source drives the enormous development of SSL market. From the mechanical point of view, the failure of the LED packaging, defined by the lumen depreciation or and color shift, is mostly multiple root-cause issues. This is because multiple materials are applied for multiple physical reactions in a single LED packaging.
In this paper, we apply the nonparametric modeling techniques, such as the k-th nearest neighborhood (kNN) method with the Fnn enhancement, and compare its prediction capability with the gate neural network. An average SPD prediction error of approximately 3-5% is observed, with 30 times shorter learning time, comparing to the pure neural network approach
| |