1. Uneven degradation and condition monitoring of multi-chip power modules for wind turbines
- Author
-
Hu, Borong
- Subjects
TJ Mechanical engineering and machinery ,TK Electrical engineering. Electronics Nuclear engineering - Abstract
The powertrain conversion system in state-of-the-art wind turbines has developed to a power rating of more than 10 MW. Due to the relatively low current rating of a single semiconductor chip, the large power module in turbine converters still adopts a multi-chip in- parallel setup, counted as the most vulnerable component in the turbine system. Thus, this thesis focuses on evaluating the uneven degradation of multi-chip power modules under realistic conditions and developing field-deployable condition monitoring methods for wind turbine converters. Two kinds of initial defects in power module solder layer, voids and cracks, indeed grow quietly under low-temperature stress cycles, illustrated by computed tomography scanning and finite element analysis. This thesis provides a physics-of-failure tool to estimate such dynamic of defect growth and finds that a void may first transform into a crack then grow more rapidly leading to device failure. At converter level, due to deep temperature cycling calculated from an electrothermal model, the machine side converters of fully and partially rated wind turbines, both consume a large amount of lifetime under the fundamental frequency. When looking inside the multi-chip module, an asymmetrical packaging layout and initial defects can cause years lifetime difference between paralleled devices while the weak one's further ageing progress will be significantly accelerated. A condition monitoring scheme for detecting such uneven degradation in a multi-chip-inparallel system is proposed in this thesis, based on a core concept - train a network to represent the healthy state and then use its prediction deviation to distinguish faulty conditions. A two-stage neural network method based on only external measurements experimentally achieves a detection rate of over 98%. Furthermore, the feasibility of such a method is improved in three aspects. The labelled data for the network training is generated from an inverter test rig of equivalently emulating uneven degradation. The fibre Bragg grating multi-point sensing technique provides high temperature measuring precision with immunity to electromagnetic interference. The complex operating conditions is also generalised by a deep neural network structure, which achieves an overall accuracy of more than 95% under dynamic thermal conditions encountered in a practical wind speed profile. Finally, based on the same concept, a field-deployable condition monitoring method is proposed to detect the early-stage fault of wind turbine converters using limited and unbalanced SCADA data. A deep neural network with optimised cost function is designed by an unsupervised approach and empowered by an online learning process for long-term real-time anomaly detection. The proposed method shows robust diagnosis results and would predict the converter fault a few days ahead of actual failure.
- Published
- 2021