Research On Pig Cough Sound Recognition Methods Based On Audio Tehnology

In recent years,Chinese pig breeding in the small-scale have gradually withdrawn from the market.The farming methods have gradually become more standardized and large-scaled.At the same time,the prevention and control of diseases in live pig houses become particularly important.Respiratory diseases are one of the major threats to the pig breeding industry.It has the characteristics of high mortality and strong infectivity,which seriously affect the health of pigs and the economic benefits of the farms.Establishing an early warning system for respiratory diseases can detect abnormal conditions in pig houses in a timely manner.Also,it can be used to give farmers a reminder in the early stages of respiratory diseases,so as to effectively deal with them in time and reduce economic losses.Cough is one of the main clinical symptoms of early respiratory disease in pigs.It is useful to utilize audio technology to identify the cough sound of pigs in the house,which can provide effective data support for the early warning system of respiratory diseases.The audio recognition technology of pig vocalizations voice recognition was used in this paper.Pig vocalizations audio database,the vocalizations feature extraction and the cough recognition methods based on convolutional neural network were established.The specific research work is as follows:(1)Audio database of pig vocalizations was established.Real-time audio data were collected in the hog house.Also,the collected data were noise mitigated to reduce the interference of periodic noise and broadband noise.The processed audio data was extracted and annotated by expert audiovisual methods.A dual threshold endpoint detection method was used to locate the starting position of the sound sample.A total of 2,744 vocalizations sound samples and 1,807 non-cough sound samples in the database were established to provide data support for subsequent research.(2)The pig vocalizations features were extracted and analysed.Power spectral density(PSD),Mel frequency cepstrum coefficient and its first-order difference coefficient(MFCC+?MFCC)and spectrogram characteristics of each audio data were extracted in the hog audio database and analyzed.It showed that the PSD characteristic parameters of coughing sound of hogs and most non-coughing sounds were obviously different within a certain frequency range.However,compared with the noncough sound,the static part MFCC of MFCC+?MFCC were significantly different,while the dynamic part ?MFCC was less.In addition,the difference between the cough sound as well as the non-cough sound of the water flow sound and the scream of the pig was the most obvious.(3)The mothod of cough recognition based on the fusion of MFCC+?MFCC features.The convolutional neural network was used to fuse the adjacent frame information in the MFCC+?MFCC combination feature to strengthen its dynamic characteristics to improve the recognition accuracy of the cough sound of the pig.The new feature generated after fusion was recorded as MFCC-CNN.A 10-fold cross-validation was used for simulation.Softmax and support vector machine(SVM)classifier were used to model 64-dimensional MFCC-CNN.A total of five performance evaluation indicators were defined to quantify the model performance,including Accuracy,Cough Accuracy,Precision,Non-cough Accuracy and F1-score.The results showed that MFCC-CNN can achieve satisfactory results.When modeling with Softmax and SVM classifiers,the 64-dimensional MFCC-CNN generated by fusing adjacent 55 frames and 45 frames of information in the combined features of MFCC+?MFCC as feature parameters could achieve the best recognition effect.It showed that the accuracy of cough sound could reach 97.81% and 97.71%,respectively.The F1 scores of the model could reach 97.25% and 97.40%,respectively.Compared with modeling using MFCC+?MFCC as the characteristic parameter,the F1-score of the model were improved by 3.85% and 5.35%,respectively.(4)The method of cough recognition based on spectrogram.Spectrogram features of each audio data were used in the pig vocalizations database as classification features.The task of recognizing cough sounds was converted into the image classification task of its spectrogram by using the advantage of convolutional neural networks in the field of image recognition.Using the Alexnet neural network,the Softmax layer was adjusted to two channels.Also,all fully connected layer parameters were retrained to be used as a classifier for different audio data spectrograms.Pig coughing sound and non-coughing sound as the unit to model,and uses 10-fold cross-validation method was also used for simulation.The experimental results showed that this method could achieve more satisfactory results.Among them,the cough sound accuracy rate and non-cough sound accuracy rate were 96.74% and 91.75%,the accuracy rate was 94.76%,and the F1-score of the model reached 95.68%.Compared with the pig cough sound established based on sequence features(specifically taking PSD as an example),its cough sound accuracy rate and model F1 score were improved by 2.04% and 2.28%,respectively.In summary,two different methods of audio technology were proposed for constructing the recognition model of pig cough.Its simulation experiments could achieve satisfactory results.The cough sound accuracy and F1-score could be reached to more than 95%.The new ideas and new methods for the automation,were provided to identify respiratory diseases precisely and intelligently.