A meta-learning approach in a cattle weight identification system for anomaly detection

Abstract

Weighing management in cattle farming is important for farmers, as it allows them to accurately monitor the growth and development of their animals. It is also a valuable tool that allows farmers to maximize the production and welfare of their animals. However, it is difficult for the farmer to detect if the herd of animals being weighed is gaining the ideal weight for a given breed and age. In addition, normally, when a new breed of cattle is introduced to a farm, there is very little data. This article proposes a meta-learning framework (MTL) for identification models used in the fattening process of animals to detect anomalies in cattle weight. The proposed MTL framework has a knowledge base of Meta-Models on Identification models based on machine learning techniques, which is used to select the identification model to use when a new breed of cattle arrives on the farm. This knowledge base is updated, either because a previous identification model has been successfully adapted to the new breed, or a new identification model has had to be generated, allowing the framework to continuously improve its performance over time. Particularly, this article presents in detail the process of adaptation of the previous identification models to new breeds carried out by our MTL framework. Besides, to test our approach, a case study is presented, using records of animals raised and fattened at the ”El Rosario” farm, located in the municipality of Monteria (Córdoba-Colombia). The results are very encouraging in terms of the ability of our framework to adapt the identification models to different possible scenarios in the process of detecting anomalous weights. In general, the identification models generated with our proposal had an of 90.8%, which suggests that the models can explain the variability observed in the data.