Abstract: The types of partial discharge faults are closely related to the severity of insulation failures in gas insulated switchgear (GIS), thus the precise identification of these faults is crucial for maintaining the stability of power supply systems. Traditional methods of partial discharge pattern recognition lack adaptive denoising capabilities and mechanisms to handle multi-scale fault features, and they primarily rely on expert experience. This results in significant limitations when dealing with complex partial discharge signals that contain substantial noise and inherent multi-scale characteristics, thereby constraining further improvements in the accuracy of fault type recognition. To address these challenges, this paper introduces a multi-scale attention adaptive denoising network (MAADNet), which integrates a multi-scale feature learning module, a convolutional block attention module (CBAM) attention mechanism module, and a soft threshold function. This network possesses robust adaptive denoising and multi-scale fault feature learning capabilities. Specifically, the multi-scale feature learning module employs dilated convolutions with varying dilation rates to extract features at multiple scales; the CBAM attention mechanism and soft threshold function collaboratively adjust the denoising threshold based on the characteristics of the input partial discharge signals, effectively suppressing noise. Moreover, to validate the effectiveness of the proposed network, a partial discharge experimental platform was set up, and four typical partial discharge fault models were designed and constructed to collect a dataset of different fault types. Experimental results demonstrate that the proposed method achieves a recognition accuracy of 94.34% on the partial discharge dataset, outperforming other advanced methods and showing promising application prospects.