Sound Recognition of Harmful Bird Species Related to Power Grid Faults Based on VGGish Transfer Learning

Bird activities threaten the safe operation of transmission lines and substations. In order to assist differentiated prevention of bird-related faults in power grid, this paper proposes a birdsong recognition method based on VGGish transfer learning. Firstly, according to the information of bird species related to historical power grid faults and the investigation results of bird species around transmission lines, 18 high-risk, 18 low-risk, and 2 harmless bird species were selected to establish a sample set with their song signals. Then, the birdsong signals were preprocessed by framing, windowing, noise reduction and clipping, thus to extract the spectrogram, which was mapped to a 64-order Mel filter banks to get the Mel spectrogram. Aiming at weak generalization ability of traditional birdsong recognition models due to insufficient number of samples, the VGGish transfer learning network pretrained by AudioSet was used as the birdsong feature extractor, and the Mel spectrograms of harmful bird species belong to the training set were taken as inputs to train the network parameters, thus to extract 128-dimensional VGGish deep features for bird recognition. This method was applied to classify 38 kinds of bird species related to power grid faults, and the recognition accuracy reaches 94.43%. The research results can provide references for power grid inspector to carry out intelligent recognition and ecological prevention of bird species.