Research On Pig Cough Sound Recognition Model Based On Mel Frequency Cepstrum Coefficients And Deep Learning

With the increasing demand for pork products in the daily lives of Chinese residents,the market demand is also gradually increasing.This has led to the gradual transformation of pig farming in China from individual farming to intensive scale farming,in order to improve the production efficiency of the pig industry.However,intensive farming increases the risk of respiratory diseases in pigs,which is one of the main reasons affecting the productivity and economic benefits of the pig farming industry.Respiratory system diseases in pigs are usually discovered by breeders during pig farm inspections.This disease warning method is not only inefficient,but also requires a large amount of labor costs.The pig cough sound signal can serve as the main basis for screening and diagnosing respiratory diseases in live pigs.Therefore,studying the pig cough sound recognition model plays a crucial role in achieving the development of intelligent intensive breeding.In order to obtain the pig sound signals required for this study,an indoor pig sound collection system was established,and the collected pig sound signals were preprocessed and Mel frequency cepstrum coefficients were extracted.Based on the principle of CNN,a pig cough sound recognition model was established.Through model training and validation,the optimal combination of pig cough sound feature parameters was analyzed and selected.Finally,the attention mechanism was utilized to improve the pig cough recognition model and further enhance its performance.The main work and conclusions include:(1)Build a pig sound collection system in the house and establish a database of pig sound signals.Design and build a raspberry pie based indoor pig sound collection system to collect pig sound signals,and use the collected pig sound signals to establish a pig sound database for subsequent research.(2)Pre processing and feature parameter extraction.Using a first-order FIR high-pass filter to pre emphasize pig sound signals;By analyzing the frequency of pig cough and environmental noise in the pigsty,the spectral subtraction method,which has a good denoising effect on additive noise,was used to denoise the pig sound signal;Using the endpoint detection method based on short-term energy to extract effective segments from pig sound signals and eliminate invalid segments from pig sound signals;Extract MFCC feature parameters,as well as MFCC first-order differential and MFCC second-order differential feature parameters,and concatenate different feature parameters to form 7 sets of feature parameter combinations that can respectively characterize the static and dynamic characteristics of pig sound signals,including 13 dimensional MFCC and 13 dimensional MFCC Δ MFCC,13 dimensional Δ 2MFCC,26-dimensional MFCC+ Δ MFCC,26 dimensional MFCC+ Δ 2MFCC,26 D Δ MFCC+ Δ2MFCC and 39 dimensional MFCC+ Δ MFCC+ Δ 2MFCC。(3)Establishment of a pig cough recognition model.Deeply explored the basic structure of CNN and the calculation and training methods of network models,and combined with CNN theory and research needs,built VGG-16,Inception V1,Res Net-34,and Dense Net-121 pig cough recognition models;By improving the Dense Block through the SENet attention module,the original recognition model is improved.The improved model is the SE-Dense Net-121 pig cough recognition model.(4)Performance testing and analysis of feature parameter combinations.Train the model with 7 sets of feature parameter combinations as inputs,and compare and analyze the performance of the feature parameter combinations based on 4 evaluation indicators.The experimental results indicate that a 26 dimensional MFCC can characterize the static and dynamic characteristics of pig sound signals+ Δ The MFCC feature parameter combination has the best performance,with a recognition accuracy of 92.9%,a recall rate of 98.6%,an accuracy rate of 96.6%,and an F1 score of 97.6%.(5)Experiment and Analysis of a Pig Cough Recognition Model Integrating Attention Mechanism.With MFCC+ Δ MFCC was used as the input training model for the model,and SE-Dense Net-121 was compared with four unimproved models for experimental analysis.The results showed that the recognition accuracy,recall,accuracy,and F1 score of the SE-Dense Net-121 pig cough recognition model were 93.8%,98.6%,97%,and 97.8%,respectively.The accuracy,precision,and F1 score were improved by 0.9%,0.4%,and 0.2%compared to before the improvement,indicating that the SENet attention module can effectively improve the performance of the model,Prove that improvement is successful.