Research on White Feather Broiler Health Monitoring Method Based on Sound Detection and Transfer Learning
Zhigang Sun,Min Zhang,Jingang Liu,Jiaqi Wang,Qianyu Wu,Guotao Wang
DOI: https://doi.org/10.1016/j.compag.2023.108319
IF: 8.3
2023-01-01
Computers and Electronics in Agriculture
Abstract:Livestock-vocalization-based health monitoring methods have the advantage of being contactless and non-invasive. In this study, the authors proposed a white feather broiler health monitoring method based on sound detection and transfer learning. The health of white feather broilers in the breeding area was determined by judging the frequency of coughs in the white feather broiler audios, and a quantitative evaluation index, cough rate, was newly proposed. Thus, the white feather broiler health monitoring problem is transformed into the multi-classification problem in machine learning, and training a reliable classification model becomes critical. First, large numbers of labelled white feather broiler audios were collected. Through signal filtering, sub-frame processing, pulse extraction and endpoint detection, the white feather broiler audios were transformed into labelled frame signals. Next, sixty sound features in multiple domains were extracted, and the labelled frame signals were converted into labelled data, thus created the data set. Then, multiple non-linear classification models were trained, the best performing classification model was performed parameter optimization, and a reliable classification model was obtained. Finally, the classification model was used to identify the sound types that appear sequentially in the unfamiliar white feather broiler audio. Based on this, the cough rate of the audio was calculated, and the health of the white feather broilers in the source breeding area was determined. In practice, there are differences in the vocalizations of white feather broilers at different days-old, and training classification models to identify white feather broiler audios of each day-old would result in a large resource loss. To address this problem, the authors considered setting white feather broilers of every three days-old as a collective, and attempted to train applicable classification models, respectively. Specifically, one of the collectives was selected, large number of labelled white feather broiler audios were collected, the master data set was created, and the conventional classification model was trained. It is based on the parameter-optimized random forest, which achieved a classification accuracy of 91.25%. For the other collectives, a small number of labelled white feather broiler audios were collected, and slave data sets were created, respectively. The improved TrAdaBoost multi-classification algorithm was used to filter instances from the master data set to each slave data set, the transfer data sets were created, and the transfer classification models were trained, respectively. They achieved a classification accuracy of approximately 83%. In addition, the definition of identification accuracy was newly proposed. Test results show that, the average classification accuracy achieved by the conventional classification model was 88.93%, and the average classification accuracy achieved by the transfer classification models was approximately 82%. This indicates the strong generalization ability of the classification models. The identification accuracies achieved by several classification models were all 100%. It demonstrates the feasibility and practicality of the method proposed in this paper.