To address the challenges of strong noise

weak fault signals

and sample scarcity in the practical application of switchgear acoustic fault recognition

and to achieve accurate and efficient online diagnosis

an improved convolutional neural network (CNN) model integrating gradient-sensitive pruning and transfer learning is proposed for switchgear acoustic fault recognition. First

to tackle the difficulty of extracting multi-scale features from acoustic signals

a multi-scale residual convolution module is designed

employing parallel dilated convolutions with different dilation rates to construct an adaptive receptive field. Second

to enhance the model's feature discrimination capability under complex noise backgrounds

an attention mechanism is introduced to build a dynamic weight allocation module for time-frequency domain features. Furthermore

to meet the requirements of real-time performance and lightweight design for on-site deployment

a gradient-sensitive structured pruning algorithm is proposed to compress the model

and transfer learning using pre-trained parameters from ImageNet is adopted to overcome the bottleneck of small-sample training. Finally

a test set is constructed using acoustic data of six typical operating states from a 110 kV substation switchgear to validate the application performance of the proposed model. The results show that

in the task of switchgear acoustic fault recognition

the proposed model achieves an average improvement in recognition accuracy of 4%-8% compared with traditional CNN

support vector machine

and other methods. After pruning

the model's inference speed increases by 18.6% with only a 0.42% loss in accuracy. The proposed model effectively balances high diagnostic accuracy with the practical requirements of on-site deployment.