Introduction  In the low-voltage distribution network, the residual current protection device, as an important guarantee of electricity safety, can reduce the harm caused by the leakage fault of electrical appliances and prevent human electrocution accidents. The current residual current protection device relies on the residual current signal size as the basis for the action of the protection mechanism, but has no function to identify the electrocution characteristics. To address this problem, this paper proposes a method for electrocution signal feature extraction and electrocution diagnosis in low-voltage distribution networks based on wavelet decomposition and denoising, as well as GRU.

  Method  In this paper, the residual currents collected from electrocution experiments were pre-processed by downsampling and discrete wavelet denoising; The time and frequency domain electrocution characteristic parameters of the residual currents were extracted by the sliding window method, and the Fourier transform was used to extract the characteristic parameters of residual currents to the second harmonic amplitude. All the extracted feature parameters were used to form a high-dimensional feature space vector; which was subject to dimensionality reduction using the method of principal component analysis to obtain a new set of three-dimensional feature vectors. A diagnostic model for electrocution was established, and the three-dimensional feature vectors representing electrocution features were input into the model. Comparison experiments were conducted on electrocution signals using five different electrocution diagnostic models, such as recurrent gated network (GRU).

  Result  The experimental results show that the convergence of the GRU-based electrocution diagnosis model is good, and the recognition rate reaches 98.33%.

  Conclusion  The method provides new insights for the research and development of a new generation of residual current protection devices and offers an effective guarantee for electrical safety.