Separation of radio-frequency signals triggered by valve switching in a HVDC converter by supervised machine learning methods

We classify, by supervised machine learning methods, the Electro-Magnetic radiation in the Ultra-High Frequency band recorded close to the transformer of one of the 12 valve branches of a 800 kV HVDC converter. In particular, we focus on the radiation triggered by the voltage stress induced by the switching between the thyristor valves in the DC-AC conversion. A multi-sensor intelligent acquisition system recorded Partial Discharge-like radio-frequency pulses. Two classification algorithms, namely an Artificial Neural Network and a Random Forest, are applied to distinguish pulses triggered by valve switching. With the chosen sensors setup, we found accuracies mostly larger than 90% with different feature combinations. In particular, a classification accuracy of 97% is obtained using only four features: phase, maximum amplitude, amplitude standard deviation and pulse power. This approach can be potentially implemented on the edge of the acquisition system to highlight the effects on the condition monitoring of insulation systems of one or few sources.