Recently, a research team led by Prof. GAO Ge and JIANG Li from Institute of Plasma Physics, Hefei Institutes of Physical Science of Chinese Academy of Sciences, studied the fault diagnosis of pulse width modulation converter and proposed a neural network fault diagnosis algorithm to solve existing problem in this field.
The results were revealed in IEEE Transactions on Power Electronics.
The pulse width modulation has the advantages of high efficiency, high power density, and high reliability. But due to the complexity of the drive systems and the diversity of fusion joint operation situations, pulse-width modulating voltage source converter (PWM-VSC) systems are prone to suffer critical failures. Therefore, research on fault diagnostic technology is of deep concern, especially on open-circuit fault diagnosis, which was what scientists have been focusing in this study.
The current fault diagnosis methods only deal with the rectification state or inverter state. Theoretical analysis showed that the fault characteristic quantities in both two states have completely different characteristics and complicated, which increases the difficulty of fault diagnosis.
In this research, when applying the proposed algorithm, researchers used only the three-phase grid side current as the characteristic quantity of fault diagnosis, and diagnosed 21 types of faults in both rectification and inverter state.
"Different from the traditional Convolutional Neural Network (CNN) architecture, a carefully crafted design allows the depth and width of the network to be increased while keeping the computing budget unchanged." Dr. DENG Xi, the first author of the paper explained the reason, "this can make better use of the computing resources inside the network."
The experimental result showed that the proposed model can accurately detect approximately 99.14% of the open-switch faults within 12.83 ms (<3/4 cycle) without an additional sensor.
It provided a foundation for the safe and stable operation of fusion power systems, and also provides reference value for other fields.
This work was supported by Comprehensive Research Facility for Fusion Technology Program of China under Contract,project funded by China Postdoctoral Science Foundation, and Natural Science Foundation of Anhui Province.
Fig.1.The architecture of the proposed model. (Image by ZHANG Li)
Fig. 2. Model accuracy of the neural network model. (Image by ZHANG Li)
