Research On Sound Recognition Of Boring Insects Based On Deep Learning

Wood boring insects are very secretive and have long-term and severe hazards.It's difficult to detect the damage in the early stage,but it spreads fast and is easy to break out.In the middle and later stages,the insect damage can be judged by the external characteristics,but it has often caused serious losses.At present,artificial observation or remote sensing images are often used to monitor wood borers in the adult stage.These methods consume lots of manpower,material and financial resources,and cannot be detected early in the occurrence of pests.Based on the above,sound recognition of boring insects based on deep learning was explored in this thesis.The models were designed separately to achieve insect recognition and noise reduction using pure and noisy pest data,and the possibility of automatic identification and early warning of pests in real environments was explored.On the one hand,four light-weight CNNs Insect Frames?1-4 were designed.Characteristics of different dimensions were extracted by adjusting the network structure,and the effects of dimension reduction with average pooling on feature extraction and insect pests recognition were compared.The log mel-spectrograms or the wavelet packet decomposition spectrums were fed into the CNNs to compare the impact of different inputs on recognition results.The results showed that,the log mel-spectrogram was more suitable for insect sound recognition.The accuracy of Insect Frames?1-4 recognition using log mel-spectrogram were more than 90%,and the average recognition time of CPU was about 0.1s-1.3s.Especially,the best model Insect Frame-2 reached 95.83%.On the other hand,a noise reduction network which named Enhancement for three types of noisy insect signals was designed,which directly learned the time-domain characteristics of the insect noise signals to achieve the predictive regression of the noise reduction insect signals.The dilated convolution was used to build the main body of the noise reduction network,which effectively increased the receptive field of the network.And the network introduced Leaky Re Lu(LRe LU)to realize the nonlinear transformation.The results showed that,the signal-noise ratio was increased by 7.02 d B after noise reduction.And when using trained Insect Frames?2 to recognize the noisy signal and the signal after noise reduction,the accuracy of the signal after noise reduction was increased by 68.75%.After retraining Insect Frames?2,the accuracy of the signal after noise reduction was further increased by 16.67%.This paper applied deep learning and sound identification technology to the automatic monitoring of larvae boring vibrations,which can improve the early warning of forestry boring insects,and has the advantages of high efficiency,simplicity and low cost.