Acoustic waveforms generated by several types of partial discharges in laboratory environment

Overview This dataset contains acoustic waveforms (effective bandwidth 25–100 kHz, ultrasonic range) generated by three types of partial discharges, in air, in a void, and corona discharge at a needle tip, together with a reference recording of ambient noise. Signals were captured in a controlled environment (anechoic chamber) using laboratory-grade measurement equipment. The dataset is intended for training and benchmarking machine learning models, particularly artificial neural networks, for the classification of partial discharge (PD) types from airborne ultrasonic acoustic emissions. Contents at a glance: 4000 waveforms total (1000 per class × 4 classes) Each waveform: 125 samples, 250 µs duration, 500 kS/s sampling rate Values stored in pascals (already calibrated from microphone voltage) Effective measurement bandwidth: 25–100 kHz File format: CSV (one waveform per row, one sample per column) Discharge sources and specimens Partial discharges were generated using three distinct configurations: Air discharge — discharges between two conductive parts separated by an air gap of 2.5 mm. Void discharge — internal partial discharges inside a 3D-printed insulator model (SLA) with an internal air void of cylindrical shape with diameter of 3 mm and 1 mm thickness. Corona discharge — generated at the tip of a needle electrode (50 um tip radius, 2.5 mm to the grounded counter-electrode). Voltage was applied using a GLASSMAN FJ Series FJ50R2.4 high-voltage source (120 W, up to 50 kV). The applied voltage for each class is encoded in the file name (e.g. PD_in_Air_12kV.csv): Measurement setup Acoustic waveforms were acquired using the following signal chain: Microphone: Brüel & Kjær 4939, 1/4″ free-field condenser microphone (usable range ~4 Hz to 100 kHz, ±2 dB up to ~80 kHz) Microphone preamplifier: Brüel & Kjær 2670; overall sensitivity (microphone+preamplifier) 4.6 mV/Pa Signal preamplifier: GRAS 12AQ at 20 dB gain, with the 20 kHz analog high-pass filter enabled Data acquisition: National Instruments PXI-5922 flexible-resolution digitizer, configured to 24-bit resolution, input range ±200 mV, sampling at 500 kS/s Acquisition trigger: Rogowski coil placed on the incoming high-voltage wire to the specimen — each waveform is triggered by the current pulse associated with a discharge event on the HV side of the test setup Post-acquisition filtering: 6th-order Bessel band-pass filter, 25–100 kHz (digital) The microphone was positioned 1 m from the discharge source in all measurements. Recordings were performed inside an anechoic chamber under stable laboratory conditions (22 °C, approx. 40 % RH). The waveforms stored in the dataset are the band-pass-filtered output; raw unfiltered signals are not included. The effective measurement bandwidth of 25–100 kHz is set by the GRAS 12AQ analog high-pass filter (20 kHz) combined with the digital Bessel band-pass (25–100 kHz). The 500 kS/s sampling rate provides a Nyquist frequency of 250 kHz, well above the upper passband edge. Triggering on the Rogowski-coil current signal ensures that each captured 250 µs window is centered on (or includes) an actual discharge event rather than random ambient noise. The reference noise file was recorded with the high-voltage source connected but no discharge activity present, so it captures the ambient acoustic noise of the chamber together with the noise floor of the full acquisition chain. Files The dataset comprises four CSV files. Each file contains 1000 rows (one waveform per row) and 125 columns (one column per time sample). Values are acoustic pressure in pascals. File Class Voltage Shape PD_in_Air_12kV.csv Air discharge 12 kV 1000 × 125 PD_in_Void_25kV.csv Void discharge 25 kV 1000 × 125 Corona_11kV.csv Corona discharge 11 kV 1000 × 125 Noise_NoPD.csv No discharge (ambient + acquisition chain) 0 kV 1000 × 125 Loading the data (Python) Minimal example — load one file and inspect its contents: import numpy as np import pandas as pd FS = 500_000 # sampling rate, 500 kS/s N_SAMPLES = 125 DURATION_US = N_SAMPLES / FS * 1e6 # 250 µs # Each row is one waveform; values are acoustic pressure in pascals. waveforms = pd.read_csv("PD_in_Air_12kV.csv", header=None).values print(waveforms.shape) # (1000, 125) # Time axis in microseconds t_us = np.arange(N_SAMPLES) / FS * 1e6 Loading all four classes into ML-ready arrays: import numpy as np import pandas as pd files = { 0: "PD_in_Air_12kV.csv", 1: "PD_in_Void_25kV.csv", 2: "Corona_11kV.csv", 3: "Noise_NoPD.csv", } X_parts, y_parts = [], [] for label, path in files.items(): data = pd.read_csv(path, header=None).values # (1000, 125) X_parts.append(data) y_parts.append(np.full(len(data), label, dtype=np.int64)) X = np.concatenate(X_parts, axis=0) # (4000, 125), pascals y = np.concatenate(y_parts, axis=0) # (4000,), class labels Suggested applications Classification of partial discharge type from acoustic signals (ANN, CNN, RNN, classical ML) Benchmarking of acoustic feature extractors (spectrograms, MFCCs, wavelets) on ultrasonic PD data Development of acoustic-based PD detection and diagnostics for high-voltage insulation systems Acoustic emission analysis in the 25–100 kHz ultrasonic band Limitations and notes The dataset reflects measurements taken at a fixed microphone-to-source distance of 1 m inside an anechoic chamber; performance of models trained on this data may differ in reverberant or noisier industrial environments. All waveforms have been band-pass filtered (25–100 kHz); information outside this band is not recoverable from the stored signals. Each class corresponds to a single applied voltage level. Voltage-dependence of the discharge signature is therefore not captured. Triggering is performed on the HV-side current signal via the Rogowski coil; the time alignment of the acoustic waveform within the 250 µs window therefore reflects the acoustic time-of-flight from the source to the microphone (~3 ms at 1 m in air at 22 °C), with the Rogowski-coil trigger providing the time reference. Acknowledgements This dataset was created as a part of the Listen2Future project. Under grant 101096884, Listen2Future is co-funded by the European Union. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the Chips Joint Undertaking. Neither the European Union nor the granting authority can be held responsible for them. The project is supported by the CHIPS JU and its members (including top-up funding by Austria, Belgium, Czech Republic, Germany, Netherlands, Norway, and Spain). The collection of the dataset was also supported by the infrastructure of RICAIP, which has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No. 857306, and from the Ministry of Education, Youth and Sports under OP RDE grant agreement No. CZ.02.1.01/0.0/0.0/17_043/0010085.