Fingerprint Recognition of Partial Discharge Signals in Deep Learning Enhanced Rydberg Atomic Sensors

TL;DR

This work employs a Rydberg atomic sensor to directly capture time-domain responses of partial discharge emissions and construct distinctive spectral fingerprints for different types, and applies a 1D ResNet deep learning model to recognize these fingerprints from time-domain signals without manual feature engineering.

Abstract

Partial discharge originates from microscopic insulation imperfections in high-voltage apparatus and is widely considered a critical marker of incipient deterioration. Conventional partial discharge detection methods are typically constrained by limited bandwidth and often rely on predefined feature extraction, which impedes reliable recognition of broadband transient signals. In this work, we employ a Rydberg atomic sensor to directly capture time-domain responses of partial discharge emissions and construct distinctive spectral fingerprints for different types. A 1D ResNet deep learning model is then applied to recognize these fingerprints from time-domain signals without manual feature engineering. Under increased source-antenna distances, where spectral features are significantly attenuated, the model attains a recognition accuracy of approximately 94\% across four partial discharge categories, demonstrating robustness to attenuation and noise. We further validate the approach in a simulated early-warning scenario, where partial discharge signals mixed with noise are analyzed and the model successfully generates predictive alarms. These results underscore the potential of integrating Rydberg-based broadband sensing with data-driven analysis for non-invasive, high-sensitivity diagnostics of electrical insulation systems.

Figures (6)

• Figure 1: Illustration of the setup and physical diagram. (a) Overview of experimental energy diagram. The probe (780nm) and coupling (480nm) laser couple the atomic states of the ground state $5\rm{S}_{1/2}$, the intermediate state $5\rm{P}_{3/2}$ and the Rydberg state $58\rm{D}_{5/2}$, forming a ladder-type electromagnetically induced transparency (EIT) scheme. In the presence of partial discharge fields, the Rydberg level $58\rm{D}_{5/2}$ exhibits an energy shift $\Delta s$ due to the ac Stark effect. (b) Experimental setup. Partial discharge signals are received by an antenna at different distances R, coupled into a waveguide, and radiated directly to the Rb vapor cell. The EIT transmission of the probe beam is detected by a photodetector (PD) to obtain spectral fingerprints. (c) Schematics of the residual neural network. The network comprises a one-dimensional convolution layer, residual blocks, a maxpooling layer and a global average pooling layer. For further details about these layers, see the "Method’’ section.
• Figure 2: Spectral fingerprint construction and recognition process. (a) Measured frequency spectrum of the void-type partial discharge signal spanning 0-6GHz. The data was obtained by directly connecting the discharge source to a spectrum analyzer to characterize the intrinsic frequency content of the source. (b) Time-domain waveform extracted from the Rydberg atom sensors when the partial discharge source is 1 cm from the antenna. (c) Saliency map visualization highlighting the model’s attention across different temporal regions of the signal. (d) Flowchart of fingerprint recognition. Partial discharge signals are transmitted via an antenna to the Rydberg atom system, and encoded as spectral fingerprints, which are then divided into training and test sets. The training set is used to train a residual network, generating a recognition model, which is then evaluated on the test set to assess recognition performance. For further details about the training and testing process, see the "Method’’ section.
• Figure 3: Recognition performance of the deep learning model across training epochs. (a) Confusion matrices of signals measured at 30 cm after 3, 8, and 100 training epochs, with overall recognition accuracies of 46.5$\%$, 71.5$\%$, and 93.5$\%$, respectively. (b-c) Evolution of loss and accuracy curves with epoch for the training set (blue) and the validation set (orange). (d) t-SNE visualization of the extracted feature representations.
• Figure 4: Performance comparison between the classical baseline and the deep learning method. (a) Confusion matrix of the FFT+SVM baseline, showing significant misclassifications between floating and particle discharges. (b) Multi-dimensional comparison radar chart of FFT+SVM (dashed line) versus 1D ResNet (solid line) across four evaluation metrics, demonstrating the comprehensive superiority of the deep learning approach.
• Figure 5: Prediction performance and early-warning of partial discharge signals. (a) Time-domain waveform of the void-type partial discharge signal measured at 30 cm. (b) Prediction accuracy versus time window length ($\Delta$t). (c) A 300 ms signal composed of void signal and noise. (d) Probability versus time for early-warning detection of the signal shown in (c), with 0.5 set as the alarm threshold. (e-f) Saliency map of the void signal and noise.