Abstract：Power cable is an important carrier of power transmission. To solve the problems of low accuracy

and large computation in voicing diagnosis with traditional convolutional neural network model，a novel cable fault

diagnosis method based on multiple voicing features and lightweight convolutional neural network was proposed.

Firstly，three feature maps were extracted，including spectrogram，mel spectrum and mel-frequency cepstral

coefficients. Secondly，a lightweight convolutional neural network model was constructed for multi-feature

extraction and fusion. Finally，the typical defect model of cable was made and verified by the voiceprint data

collected on site. The verification results show that the proposed method can achieve reliable diagnosis of three

typical fault types with an average accuracy of 97.6%. Compared with only using a single spectrogram，mel

spectrum or mel-frequency cepstral coefficients feature diagnosis method，the accuracy of this method was

improved by about 1.3%，3.1% and 3.1% respectively. Meanwhile，compared with AlexNet，MobileNetV2 and

GoogleNet，the accuracy of the proposed method was increased by 1.3%，0.7% and 4.4%，the training time of each batch was reduced by 172.3 s，18.2 s and 69.4 s，and the model file size was reduced by 193.4 MB，14.6 MB and 25.7 MB respectively.

      Key words：power cable；voice print；convolutional neural network（CNN）；partial discharge；multi-feature

fusion