Abstract
Leakage current is a critical parameter for monitoring the operational health of distribution network surge arresters. While polymagnetic current sensors offer miniaturization and high accuracy, their performance in the field is often degraded by complex electromagnetic interference (EMI). This paper proposes a novel leakage current measurement method using a hybrid pipeline of Ensemble Empirical Mode Decomposition (EEMD), Singular Value Decomposition (SVD), and a low-rank Radial Basis Function (RBF) neural network. The method decomposes the signal to isolate eigenfrequencies, extracts non-eigenfrequency interference, and utilizes a low-rank RBF network to suppress white noise. Experimental results demonstrate a signal-to-noise ratio (SNR) improvement of approximately 20dB and a maximum average absolute error of only 2.62%, enabling reliable health diagnostics in high-noise environments.
Introduction
Surge arresters are essential components for protecting power grid infrastructure from lightning and switching surges. As these devices age or succumb to moisture ingress, they exhibit minute leakage currents (microamperes to milliamperes). Detecting these currents is the primary method for predictive maintenance.
Modern monitoring solutions increasingly rely on polymagnetic sensors based on the Tunnel Magnetoresistance (TMR) effect due to their high sensitivity and small footprint. However, these sensors are highly susceptible to external electromagnetic interference and internal circuit noise. In real-world substations, this interference often masks the underlying 50Hz leakage signal, necessitating advanced digital signal processing to recover usable data.
Proposed Denoising Pipeline
The authors introduce a multi-stage "de-disturbance" strategy designed to target specific noise types:
- EEMD (Ensemble Empirical Mode Decomposition): Standard EMD often suffers from "mode confusion" when signal frequencies overlap. EEMD addresses this by adding controlled white noise to the signal, facilitating a more stable decomposition into distinct Intrinsic Mode Functions (IMFs).
- SVD (Singular Value Decomposition): The resulting IMFs are processed via SVD. By analyzing singular values, the algorithm identifies and removes characteristic interference frequencies—such as those generated by nearby switching equipment—while preserving the fundamental 50Hz component.
- Low-Rank RBF Neural Network: This stage targets random white noise. By intentionally restricting the "rank" (the number of hidden layer neurons) of the RBF network, the model is forced to learn only the underlying signal trend. Because white noise is high-frequency and random, it cannot be effectively mapped by a low-rank network, resulting in a naturally smoothed output.
- Gaussian Window Function: A final pass ensures waveform smoothness, preparing the signal for diagnostic analysis.
Experimental Evaluation
The method was validated using both simulated data and a physical experimental platform. The simulation superimposed a 500uA 50Hz sinusoid with four characteristic interference frequencies (800Hz to 3100Hz) and random white noise.
The physical platform utilized a TMR-based polymagnetic current sensor integrated with a microprocessor and a signal generation system to test input currents ranging from 50uA to 1500uA. This setup allowed the researchers to simulate the "dirty" electrical environments typical of distribution networks.
Key Results and Discussion
The EEMD-SVD-RBF method was benchmarked against several standard techniques, including EMD-SVD and STFT-SVD. Key findings include:
- SNR Enhancement: The proposed method achieved an SNR improvement of ~20dB across all test cases.
- High Fidelity: For extremely weak 50uA signals, where noise dominance is highest, the Normalized Correlation Coefficient (NCC) reached 0.952, indicating near-perfect reconstruction of the original waveform.
- Precision: The maximum average absolute error was restricted to 2.62%, ensuring that the measured leakage current accurately reflects the arrester's state.
The "low-rank" approach proved particularly effective; by treating a neural network's structural limitation as a filtering feature, the authors successfully suppressed noise that traditional filters often leave behind.
Conclusion
This research provides a robust framework for enhancing the reliability of TMR-based sensors in power grid monitoring. While the computational demands of EEMD and RBF training may require optimization for low-power field devices, the significant gains in signal clarity represent a major step forward for automated grid maintenance. Future work will focus on the system's performance during transient surge events and long-term stability under changing environmental conditions.
Summary updated