Research On Pig State Recognition Based On Audio Feature Fusion

Breeders are always faced with emergencies such as pig cough and stress response during the pig breeding management.It is helpful to master the health and physiological status of pigs in time for breeders by monitoring the status of pigs in real time.Therefore,the identification of pig status is an important part of pig welfare breeding.This study uses speech recognition technology to recognize pig audio in order to monitor the state of pigs.The main research contents of this thesis include the preprocessing of pig audio signal,the extraction and fusion of pig audio features,the construction of pig audio recognition model based on Bi-LSTM,and design and implementation of pig audio recognition system.The specific research work is as follows:1)Audio database of landrace pig vocalization was established.In order to obtain a signal and pure pig sound corpus,the continuous pig audio is preprocessed by multi window spectral estimation,spectral subtraction denoising,endpoint detection and so on.An audio database containing 3750 sounds of eating,estrus,cough,howling and humming of Landrace pigs was established by marking a single pig sound corpus manually.2)The pig audio features were extracted and fused.The MFCC and LPCC of each audio in the audio database were extracted.In order to improve the representation ability of audio features to pig vocalizations,MFCC and LPCC are combined and fused to obtain FUSION features and FUSION(Fisher)features respectively.3)The pig vocalizations recognition model was established.In order to recognize feeding,oestrus,coughing,howling and humming of pigs effectively,pig vocalizations feature datasets were made by pig audio features and one-hot codes,and acoustic model of pig vocalization recognition based on Bi LSTM was established.The acoustic model was used to model four audio features(MFCC、LPCC、FUSION and FUSION(Fisher))respectively and the accuracy was defined to quantify the model performance.Through the 5-fold cross-validation experiment and analysis,the average accuracy of model could reach 89.79%,84.80%,92.35% and 94.13% respectively.The feature fusion method based on Fisher criterion is adopted,when 32 dimensional fusion(Fisher)feature is selected for modeling,the model could achieve the best recognition effect.Through the 5-fold cross-validation experiment and analysis,the average accuracy of model was 96.48%.The algorithm application test was carried out with noisy pig sound corpus,compared with modeling with signal feature(MFCC,LPCC),modeling with fusion feature(FUSION,FUSION(Fisher))had higher accuracy in the low SNR environment.4)The pig vocalizations recognition system was designed and implemented.The Nano PC-T4 development board was used to collect and transmit pig vocalization data.The server processed the received audio data,and realized five kinds of pig vocalization recognition by using the established acoustic model.