Neural network-based fault diagnosis for electric power meters

Intelligent algorithms are increasingly applied in the smart monitoring of power systems, demonstrating significant advantages particularly in fault diagnosis tasks for power metering equipment. Addressing the limitations of traditional methods in processing high-dimensional non-stationary data, this study constructs an intelligent fault diagnosis model for power metering based on deep neural networks. The model employs Short-Time Fourier Transform (STFT) to extract time-frequency features from raw voltage and current waveforms, constructing a three-channel spectrogram input. Convolutional Neural Networks (CNN) are used for local feature encoding, combined with Bidirectional Gated Recurrent Units (BiGRU) to enhance temporal modeling capabilities. Additionally, activation heatmaps and KL divergence regularization terms are introduced to improve diagnostic interpretability and discrimination ability. Experiments conducted on 2.64 million real-world data points covering 12 typical metering fault categories demonstrate that the proposed model achieves 96.78% accuracy, 95.42% recall, and 95.93% F1-score. These metrics outperform traditional methods such as Support Vector Machine (SVM) and k-Nearest Neighbors (KNN) while maintaining strong robustness in high-noise environments. This research provides an efficient and scalable deep learning solution for intelligent fault identification in power systems.