Abstract
This study proposes a novel deep learning framework for anomaly detection in coal mine hydraulic support pressure data. The framework innovatively combines bidirectional LSTM and CNN architectures, incorporating gated residual connections and self-attention mechanisms to effectively capture both temporal features and local patterns in pressure data. The research utilizes pressure data collected from 10 hydraulic supports on a fully mechanized mining face in western China, spanning from July 1st to July 30th, 2022. Through a systematic data preprocessing pipeline, including temporal resampling, missing value handling, normalization, and other steps, a standardized experimental dataset was constructed. Experimental results demonstrate the model's excellent performance in anomaly detection tasks. Ablation studies validate the effectiveness of key components, with CNN layers and gated residual mechanisms contributing most significantly to model performance, leading to test loss increases of 472% and 352% respectively when removed. The model exhibits strong capabilities in sample reconstruction and anomaly identification, effectively distinguishing between normal pressure variations and anomalous patterns. However, the method has certain limitations, including dependency on data quality, fixed threshold strategy constraints, and computational complexity issues. These findings provide important references for improving hydraulic support monitoring systems and have practical significance for enhancing coal mine safety production.