The Feature Extraction Of Fruit Fly's Sound And The Research Of Artificial Neural Network Classification

Insects are closely related to human beings. They have direct or indirect influence on people's life. Insects communicate through particular sounds produced by various actions. The sounds produced by different actions mean differently. Studying the insect sound, analyzing its causes and features, and making comparison between different groups are of great significance to learn the rules of insects' actions and monitor and control the pests. The classification by sounds is helpful to classify the related species, dubious species and subspecies.This paper has collected and analyzed the fruit fly's sound.In the experiment, all the fruit flies are drosophila melanogasters provided by the School of Life Science of Shaanxi Normal University. There are three strains of them. One is Canton-Special(Canton-S) and the other two are mutant fruit flies numbered 18 and 22 respectively. Before the collection of sounds, each type is divided artificially into the male and the female. The wing-vibrating sounds of the male and the female are collected separately. The sounds are recorded and saved in WAV form without compression and aberration.Firstly, the wing-vibrating sound signals of fruit flies, which have noise, are processed by means of wavelet denoising and adaptive filtering. What the results indicate are as follows. It is effective in controlling the noise in high frequency part using wavelet denoising. The energy of noise is in the low frequency part, around 50Hz in the spectrum. The noise in the low frequency is not eliminated effectively by wavelet denoising. Therefore, adaptive filtering proves to be an effective approach of removing noise in this research.Secondly, the male and the female fruit flies of the three strains have been analyzed in terms of sounds from the time domain and frequency domain. With the analysis of time domain, the wing-vibrating sounds of the three strains have similar sine wave with slight difference in the interpulse interval (IPI). With the analysis of frequency domain, the male and the female of the same strain have a large overlapping part in the frequency of wing-vibrating sounds. On average, the frequency of the male is a bit higher than that of the female. Different strains also have overlaps in the frequency of wing-vibrating sounds, and little difference. The main frequency of the three strains ranges from 200 to 300Hz, frequency from 0 to 4000Hz, energy from 200 to 2000Hz. Number 22 has the largest range of energy from 200 to 3000Hz. There are many frequency components. It presents a multi-harmonic spectrum.Thirdly, the classification of the three strains of fruit fly is conducted using neural network. The experiment is carried out in two ways, each of which is divided into four groups. Firstly, the spectrum analysis of each sound sample is made to take seven frequencies with the largest amplitude as the eigenvalue. The neural network is set in two implicit strata. As the result indicates, the neural network doesn't work effectively in classifying the male and the female of same strain and classifying the strains of same species. Secondly, the spectrum analysis of each sound sample is made to take 129 points in power spectral density estimate as the feature vector. Sound signals of the male and the female of each strain are divided into training collection and validation collection. Each training collection contains 50 signals which are used to train the artificial neural network. After the network is successfully trained, the neural network will be validated by the 30 corresponding validation samples of each strain of fruit fly. Neural network has a high level of identification accuracy for female Canton-S as well as for male number 18 and female number 22; while it has a level of identification accuracy for the other sex. It is clear that the male and the female of same strain have basically the same pattern of wing-vibrating sound, wing-vibrating frequency and sound intensity as well. The sex classification of same strain can not be conducted by sound. When the female fruit flies of Canton-S, number 18 and number 22 are analyzed, the average accuracy of identification of Canton-S reaches 96.7%, no.18 is 100%, and no.22 is 100%. As the result of the experiment demonstrates, the established neural network is effective in identifying the wing-vibrating sounds of different strains of same species. The utilization of sound signal features is applicable and feasible with great significance being popularized suitably. The features of frequency domain of wing-vibrating sounds have important significance to identification.