| Pig audio contains rich information,which can reflect the health of pigs.Completing real-time monitoring of pig audio through computer related technology can help keepers to find anomalies in time and promote pig welfare farming.Digital signal processing and audio recognition technology were used to realize the identification of live pig audio and the monitoring of pig status.The main contents included preprocessing of live pig audio,simulation of underdetermined blind source separation scenario,construction of audio recognition model and design and implementation of pig audio recognition platform,specifically:1)Audio signal pretreatment.Kalman filter was used to filter and denoise the audio signals collected from pigs.The effective segment of pig audio was selected by the double-threshold endpoint detection algorithm based on short-time energy and short-time zero crossing rate.2)Simulation test of underdetermined blind source separation of pig audio.Based on signal sparsity,affinity propagation algorithm and minimum l_pnorm,the underdetermined blind source separation experiment of pigs was completed by using the audio signals obtained from different coefficients mixed with the calls of landrace pig in different states as the observation signals.The results showed that,in underdetermined blind source separation of pigs with 3 source signals and 2 observation signals,the similarity coefficient,signal-to-noise ratio and mean square error of the separated audio signals and the corresponding source signals at different time ranged from 0.67～0.92,7.9～9.7d B and0.005～0.08,respectively.The separation performance of the algorithm had nothing to do with the length of time and the number of tests.The results had a certain stability.3)Construction of pig audio recognition model based on DNN-HMM.The deep neural network and hidden Markov model theory were used as the basis for the audio signal recognition of live pigs.Feeding sound,estrus sound,howling,hum of landrace pigs and pant of sick landrace pigs were used as the identification objects.Taking the extracted 39-dimensional Mel-frequency cepstral coefficient as the data set of network learning and recognition,a live pig state audio recognition model based on deep neural network and hidden Markov model was constructed.The experimental results showed that,in the deep neural network-hidden Markov model with 5 hidden states and 128 nodes in 3 hidden layers,for 5 kinds of live pig state audio,namely.The recall rate,precision and specificity of feeding sound,howl,hum,estrus sound of landrace pigs and pant of sick landrace pigs were no less than 73.7%,70%and 92.6%,respectively.Compared with the traditional Gaussian mixture model-hidden Markov model,the average recall rate,accuracy and specificity of DNN-HMM were 12.42%,17%and 4.14%higher,respectively.4)Construction of pig audio recognition model based on CNN.The convolutional neural network theory was taken as the basis for the audio signal recognition of pigs.Feeding sound,estrus sound,howl,hum of landrace pigs and pant of sick landrace pigs were taken as the objects of recognition,and the extracted spectrograph features were taken as the data set for network learning and recognition.A pig state audio recognition model based on CNN was constructed.The results showed that,under the convolutional neural network with 2 convolutional layers,2 pooling layers and 1 fully connected layer.The recall rate,accuracy and specificity detection of each audio recognition of live pigs were not less than94%,94%and 98.8%,respectively.The established model had a higher accuracy.5)Design and implementation of pig audio recognition platform.The Nano PC-T4development board was used to complete the collection and transmission of pig audio,and multiple functional pages were designed in combination with HTML technology to realize the recognition of pig audio using the CNN model with a trained and higher recognition rate and simulate underdetermined blind source separation of pig audio online. |
PDF Full Text Request