Abstract
Rectifier transformers are key components of high-frequency power supplies; their waveforms play an important role in diagnosing faults of high-frequency power supplies. Typically, the integral value of the current waveform is analyzed to diagnose rectifier transformer faults. However, the waveforms of currents are greatly influenced by the load, which results in serious fault-coupling problems. Generally, conventional methods cannot accurately locate faults, and they have high error rates. In this study, primary and secondary waveform currents were analyzed to extract fault feature data. The extracted data were used to train least-square support vector machines to build fault classifiers, thus realizing the fault detection and isolating the rectifier transformer. The proposed method was applied to actual waveform data from the Baosteel power plant; it performed satisfactorily.
Highlights
Metallurgy, petroleum, chemical, cement, and other industries, high-frequency power supplies are widely used for electro-static precipitators (ESPs) to improve dust removal efficiency.[1]
Extraction of current waveform features. This approach takes the view that some time-domain characteristic variables, like the amplitude ratio of adjacent waveforms,[13] cycle time, before/after zero point time ratio,[14] absolute integral value ratio of adjacent waveforms,[6] and waveform complexity,[15] should be extracted to enable accurate, efficient descriptions of current waveforms
The fault classifiers were constructed by combining least-square support vector machine (LS-SVM) and 1-v-1 multi-class expansion
Summary
Metallurgy, petroleum, chemical, cement, and other industries, high-frequency power supplies are widely used for electro-static precipitators (ESPs) to improve dust removal efficiency.[1]. An advanced approach is proposed to diagnose the faults for the rectifier transformers of high-frequency power supplies It first extracts fault features from the system’s primary and secondary currents with a designed feature extraction method, and it constructs least-square support vector machine (LS-SVM) classifiers using featured samples from different operating statuses. A more intuitive method of feature extraction was proposed for the diagnostic problem of the system described in this article This approach takes the view that some time-domain characteristic variables, like the amplitude ratio of adjacent waveforms,[13] cycle time, before/after zero point time ratio,[14] absolute integral value ratio of adjacent waveforms,[6] and waveform complexity,[15] should be extracted to enable accurate, efficient descriptions of current waveforms.
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call