Research And Application Of Bird Sound Recognition Based On Deep Learning

In actual transmission process,transmission lines are always faulty and damaged due to bird-caused damages such as nest,bird droppings,bird shorting and bird pecking.Therefore,detecting birds around transmission lines and driving them away in time is one of the main contents of transmission line operation and maintenance in the power industry.At present,most of the anti-bird devices on transmission lines rely on image recognition.However,traditional image recognition models require a lot of computing resources,which makes embedded devices such as anti-bird device unable to undertake the long-term detection tasks.Therefore,this type of model cannot exert the best detection ability,and it is impossible to prevent bird-caused damages of transmission line reliably and efficiently.This thesis introduces popular deep learning in the task of detecting bird sounds around transmission lines.In this context,this thesis carries out the birds sound recognition method based on deep learning.Firstly,it analyzes the characteristics of bird sound and bird-caused damage on the transmission lines,and selects the LogMel diagram as the input features.Secondly,it introduces DNN,CNN,RNN and CRNN models,and builds the detection models based on each model.Experiments prove that the CRNN-based detection model has the best performance,and this model is taken as the baseline model.Then,starting from improving detection speed and reducing the size of model,this thesis deeply studies the lightweight network,MobileNet.To form a detection model based on MobileNet-RNN,it improves the small version of MobileNetv3 and use this to replace the CNN part of the baseline model.The experimental data shows that the model achieves a good balance between the detection accuracy and the calculation speed.This model has a small number of parameters,which can be used in embedded devices such as anti-bird device.Finally,this thesis adds a sound detection module to the traditional bird repellent device based on image recognition,and applies the MobileNet-RNN-based bird sound detection model.By assembling the hardware and developing software,a new type of anti-bird device is formed.It is based on sound detection and assisted by image recognition.The results of laboratory tests and field tests show that the new anti-bird device can accurately and quickly detect and judge the birds around the transmission line,and realize timely and effective repelling.The overall performance of the new device is higher than the traditional anti-bird device based on image recognition.