Fault diagnosis of coal mine power grid based on hybrid deep learning

The conventional fault diagnosis technologies for coal mine power grids, including manual inspection, protection-criterion-based methods and statistical-signal-processing-based methods, suffer from poor real-time performance, limited adaptability to variable-load conditions and susceptibility of diagnostic results to noise interference. In addition, deep-learning-based power grid fault diagnosis techniques suffer from low feature extraction accuracy when applied underground and are difficult to adapt to edge devices with limited computing resources. To address these issues, a coal mine power grid fault diagnosis model, MSCA-BiGRU, was developed by integrating a multiscale one-dimensional Convolutional Neural Network (CNN), a Squeeze-and-Excitation (SE) channel attention mechanism and a Bidirectional Gated Recurrent Unit (Bi-GRU). The model used parallel one-dimensional convolution branches with kernel sizes of 1, 2, and 3 to extract multiscale features, adaptively recalibrated the fused features through the SE channel attention module, and used Bi-GRU to extract bidirectional correlation information from the feature sequences and encode the features. A coal mine power grid fault simulation model was built in Matlab/Simulink, and a fault dataset containing normal operation, single-phase-to-ground, two-phase-to-ground, two-phase short-circuit, three-phase short-circuit, and three-phase-to-ground conditions was constructed to validate the MSCA-BiGRU model. The results showed that the model achieved accuracy, precision, recall and F₁ score of 85.19%, 85.21%, 85.00%, and 84.97%, respectively, with 85 278 parameters. The model achieved a good balance between diagnostic performance and model size and outperformed the CNN, CNN-BiLSTM, and CNN-BiGRU comparative models.