Pig Cough Sounds Recognition Based On Deep Learning

In recent years,the state has focused on transforming the economic development mode of the hog industry and guiding the development of the hog industry in the direction of scale,intensification and standardization.At the same time as the largescale development of the pig industry,respiratory diseases have become one of the most common and most serious diseases in major pig farms.Pig cough is the main symptom of respiratory diseases in pigs,especially early symptoms.Therefore,early warning of early respiratory diseases in pigs can be carried out by monitoring the cough of pigs.The rise of deep learning has made fundamental changes in the processing methods of speech and image,which greatly promoted the development of artificial intelligence.In this paper,the deep learning method applied to human voice recognition is applied to the field of pig cough sound recognition.The speech recognition technology is used to analyze the cough sound of pigs and non-pig coughing sounds.The cough sound of pigs is recognized,which is helpful for early warning of respiratory diseases in pigs and promotes pigs.The development of healthy farming.In the process of this research,the environmental sound signal of the farm was collected,and the voice signal was processed by the speech signal processing method.The feature parameters were extracted,and the coughing sound of pigs was identified based on the isolated words and continuous speech recognition technology.The main contents and conclusions include:(1)Constructed an isolated word pig cough sound recognition system and a continuous pig cough sound recognition system corpus.In the isolated word cough recognition system,the original 1440 pig sound samples were obtained by intercepting and manually marking,and then the final corpus was obtained by speech enhancement and endpoint detection.The pig cough was analyzed first in the continuous pig cough recognition system.The time domain characteristics of the sound,then construct a threshold range based on this feature,obtain 222 experimental corpora by means of the endpoint detection algorithm,and perform manual sentence level annotation to construct a corpus;(2)The pig sound characteristic parameters were extracted.In the isolated word recognition system,1440 pig sound samples are time-scaled by time warping algorithm,and then 300-dimensional short-time energy feature parameters and 720-dimensional Mel-frequency cepstral coefficient characteristic parameters are extracted;in the continuous pig cough sound recognition system Extracting 26 Weimer frequency cepstral coefficient characteristic parameters of each frame signal;(3)Construct an isolated word pig cough sound recognition system based on deep belief network.The deep belief network model structure is set by empirical formula,and the model unsupervised pre-training is carried out by using the contrast divergence algorithm,and then the error backpropagation algorithm is used to carry out the model fine-tuning.In the test process,the best group pig cough recognition rate was 94.12%,the false recognition rate was 7.45%,and the total recognition rate was 93.21%.Further,the principal component analysis method was used to extract the 98.01% principal component of the 1020-dimensional pig sound sample characteristic parameter,and the 479-dimensional characteristic parameter was obtained.Through testing,the recognition rate of the best group pig cough increased by 1.68 percentage points,the false recognition rate decreased by 0.62 percentage points,and the total recognition rate increased by 1.08 percentage points;(4)Construct a continuous pig coughing sound recognition system based on twoway long-term memory network-connected time series acoustic model.Using the "pig coughing sound" and "non-pig coughing sound" as the acoustic model modeling unit,the continuous law of pig continuous sound was learned through the two-way long-term memory network,and the end-to-end pig continuous coughing sound recognition system was realized by the connection timing classification..The model is trained by time-based error back propagation algorithm.The test process uses the prefix beam acquisition and decoding algorithm to decode.The recognition rate of pig cough of the best group was 93.63%,the false recognition rate was 1.35%,and the total recognition rate was 95.24%.At the same time,the algorithm was applied and tested with 1 h corpus outside the data set.The recognition rate of pig cough was 94.23%,the false recognition rate was 9.09%,and the total recognition rate was 93.24%.