Convolutional Neural Networks and Thresholding Approaches for Single and Multi-Sensor Detection of Partial Discharges in Electrical Machine Stators

The repetitive occurrence of partial discharges (PDs) in electrical machines degenerates the insulation system and causes a premature breakdown. Consequently, PD tests at the end of the production line in accordance with the standard IEC/TS 61934 are mandatory for automotive hairpin stators to ensure that the repetitive PD inception voltage (RPDIV) meets the minimum requirements for a PD-free operation over the product lifetime. However, the test results strongly depend on the sensitivity of the utilized sensor and the PD detection algorithm. This paper investigates various single and multi-sensor-based approaches to increase the PD detection rate while ensuring robustness to noise interference. Different convolutional neural network (CNN) architectures are trained to classify between PD and noise. A profound case study for rating the detection methods is designed utilizing artificial multi-channel test signals closely representing real measurements. The CNN models outperform the conventional thresholding approach based on bandpassfiltered signals, particularly under low signal-to-noise conditions and for noise with pulse-shaped interferences. A sensor fusion with a logical OR operation increases the PD detection rate compared to the single-sensor evaluations of the Horn antennas and the high-frequency current transformer. All classifiers are successfully validated under different noise conditions and for stators not incorporated into the training process. A final RPDIV comparison with the state-of-the-art PD tester on real-world data verifies an increased detection sensitivity of the CNNs and the proposed fusion concepts, in particular for stators with small PD pulses around the inception level.