Seq2Seq-based GRU autoencoder for anomaly detection and failure identification in coal mining hydraulic support systems.

In coal mining operations, the stable operation of hydraulic supports is crucial for ensuring mine safety. However, the nonlinear, non-stationary characteristics and noise interference in hydraulic support pressure data pose significant challenges for anomaly detection and fault diagnosis. This study proposes an anomaly detection and failure identification method based on Gated Recurrent Unit Autoencoder (GRU-AE), aimed at achieving anomaly detection in hydraulic support pressure data and equipment failure early warning. Through in-depth analysis of data from two coal mines in China, we systematically evaluated the model's key parameters. The study revealed that window size had a limited impact on model performance, with a window length of 144 demonstrating optimal comprehensive performance in both anomaly detection and failure mode identification. The study also investigated the effectiveness of teacher forcing techniques. Although this technique can accelerate model convergence, it may lead to training instability and reduced generalization capability, requiring careful consideration in practical applications. Our proposed Recurrent Reconstruction Network model demonstrated excellent performance in complex coal mine hydraulic support data, effectively identifying anomalous regions and potential equipment failure characteristics while revealing potential deviations between model predictions and actual data, demonstrating its superior learning capability for periodic data patterns and equipment failure characteristics. Experimental results validated the effectiveness of the GRU-AE model in hydraulic support pressure anomaly detection and equipment fault diagnosis, providing an innovative technical solution for coal mine safety monitoring.