In the valve cooling system of converter station, the main circulation pump monitoring method based on fixed threshold is easy to ignore the coupling changes between pressure, flow, motor current, vibration and temperature signals. This paper proposes an improved graph Convolutional neural Network (IGCN) for pump fault detection under cooling load changes. 38640 synchronous operation records were collected from the two converter stations, covering normal operation, flow attenuation, pressure fluctuation, bearing vibration, seal leakage, and motor current abnormalities. Each sensor channel is represented as a graph node, and adaptive edge weights are calculated based on operational correlation, device connectivity, and fault response delay. The temporal residual aggregation was embedded into the graph convolution propagation process to retain the short-term fluctuation pattern. The dataset is divided into training, validation and test sets at 8:1:1. Experimental results show that IGCN achieves 96.1% accuracy, 94.8% recall rate and 95.6% F1 value, and the average inference delay is 38 ms, which supports stable online fault detection applications of valval-cooled pump.
