| With the development of intensive and large-scale poultry breeding industry,the poultry breeding industry and agricultural engineering researchers attach more importance to the research of using information monitoring technology to replace the traditional manual inspection in the practice of monitoring the physiological characteristics and behavioral activities for poultry.Poultry vocalization contains a lot of information,and part of the physiological state and behavioral activities of poultry can be understood by mining the information contained in poultry vocalization.Therefore,based on the audio analysis technology,this paper carried out the research on recognizing poultry vocalization in the conditions of avian influenza infection and eating.Two groups of specific pathogen free white leghorn poultry were raised in the animal breeding room of College of Veterinary Medicine,South China Agricultural University.In the first group,the H7N9 avian influenza virus extracted from the Laboratory of Animal and Biosafety Lever Three was used to infect the poultry by dripping to their noses and eyes,and audio of poultry infected by avian influenza was collected.In the second group,by feeding the poultry,we recorded the eating audio when they are eating.The audio signals of poultry were preprocessed,which includes pre-emphasis and frame by windowing.The high-pass filter was used to pre-emphasize the audio of the poultry,so as to highlight the vocalizations and weaken the environmental noise.Then,the characteristics of rectangular window,hamming window,hanning window and triangular window were analyzed,and the rectangular window and hamming window were selected to add frame the audio after pre-emphasize,so as to obtain the quasi-steady audio segment.The analysis of the poultry audio includes time-domain analysis and frequency-domain analysis.Time-domain analysis includes short-time zero crossing rate analysis,short-time energy analysis and autocorrelation function analysis,and frequency-domain analysis includes frequency-domain component analysis and spectral entropy analysis.Based on the results of time-domain analysis and frequency-domain analysis,three kinds of automatic interception for poultry vocalization are proposed: the interception method based on the short term zero crossing rate and short term energy,the interception method based on the maximum autocorrelation function,and the interception method based on the spectral entropy.The automatic interception method is used to extract the audio clip of the vocalization of the poultry group,which is convenient for the follow-up research on the recognition of the poultry vocalization.In the research of the vocalization recognition of avian influenza poultry,two recognition methods were proposed: Method for vocalization recognition of avian influenza poultry based on T-S fuzzy neural network and time-domain characteristics;Method for vocalization recognition of avian influenza poultry based on support vector machine and Mel frequency cepstrum coefficient.The first method firstly extracted the vocalization of poultry by the interception method based on the short term zero crossing rate and short term energy,then constructed two bell membership functions of T-S fuzzy neural network to recognize the audio features which extracted from the poultry vocalization.After training,the accuracy of T-S fuzzy neural network for recognizing the time domain features of avian influenza poultry vocalization in test ranges from 75% to 80%.The second method firstly extracted the 12-dimension Mel frequency cepstrum coefficient as the feature of poultry vocalization,then constructed three kinds of support vector machine with polynomial kernel function,radial basis kernel function and S-shape kernel function.The optimal kernel function parameter g and the optimal penalty parameter C of each support vector machine with different kernel function were selected on the validation set by using the4-mean split cross validation method.After training these three support vector machine,the accuracy for recognizing avian influenza poultry vocalization ranges from 87% to 90%(polynomial kernel function),86.5% and 89.5%(radial basis kernel function),84% and90%(S-shape kernel function).In the research of the vocalization recognition of poultry during eating,two recognition methods were also proposed: Method for vocalization recognition of poultry during eating based on time sequence model;Method for vocalization recognition of poultry during eating base on lightweight two stream neural network.The first method construct the poultry vocalization recognition model set PV-net by using the recurrent neural network,long and short term memory neural network and gate recurrent unit,which have the advantage of time sequence processing ability.After trained,PV-net get the accuracy for recognition of poultry’s eating vocalization between 93.5% and 96%.In the second method,based on the different data processing advantages of convolutional neural network and time sequence model,the teacher model and student model of convolutional neural network and time sequence model were constructed respectively.Using the distillation training method,the student models of convolutional neural network and time sequence model were trained by their trained teacher models respectively.Finally,the weighted fusion of two student models trained by the teacher model into a lightweight two stream neural network can get the accuracy up to 97% on the poultry eating vocalization.In this paper,methods for recognition of the vocalization of avian influenza poultry and poultry eating vocalization were proposed,and the methods were tested by the experiments operated in the animal breeding room of college of veterinary medicine of South China Agricultural University.The results show that these methods can effectively recognize the avian influenza poultry vocalization and the eating poultry vocalization,which provide a new way for research of monitoring physiological state and behavior of poultry in the future. |
PDF Full Text Request