Early detection of anomalies in cattle growth is essential for ensuring animal welfare, improving productivity, and enabling timely management interventions in livestock farming. However, centralized machine learning solutions raise concerns about data privacy, ownership, and heterogeneity across farms. This paper proposes a federated learning (FL) framework for detecting anomalies in beef cattle fattening, which enables decentralized model training while preserving data confidentiality. The system employs deep neural networks trained locally on real and synthetic farm datasets, and aggregated via the Federated Averaging function. Two experimental scenarios were designed to evaluate the system under different variance conditions: (i) balanced variance across farms, and (ii) high variance concentrated in one farm. In Scenario 1, the global model achieved a Mean Absolute Error (MAE) of 2.53 kg and a coefficient of determination (R2) of 0.9974. In Scenario 2, where a single farm introduced significant variability, the global MAE increased to 2.96 kg, with an R2of 0.9884. The anomaly detection mechanism, based on a residual thresholding method using the 2σ rule, identified 399 anomalies 4.87% in the balanced scenario. Results demonstrate the feasibility of using FL in livestock settings for robust and privacy-preserving anomaly detection, and highlight the need for adaptive techniques to address inter-farm variability.

Feature engineering plays a main role in the construction of machine learning models. In particular, it consists of several processes, including feature selection, fusion, and generation. Specifically, feature selection is the process of choosing the most relevant features from a dataset to train a machine learning model, with the goal of improving model performance and reducing training time and model complexity.. In this work, we propose a new method of feature selection based on the LAMDA (Learning Algorithm for Multivariate Data Analysis) algorithm for the classification and clustering contexts. LAMDA calculates the membership degree of an individual to a class or cluster, considering the contribution of all its features/descriptors. Our approach uses this idea to determine the most relevant features of a class (with the highest membership degree), and from there, select those with the highest membership degree across all classes. For the classification, we use the LAMDA-HAD algorithm, and in the clustering, we use the LAMDA-RD algorithm. We have carried out different experiments with different datasets, and we have obtained satisfactory results. We observe that the features selected with our proposal have a significant impact on the performance of the classification and clustering models.

Explainability analysis is a very relevant topic today, due to the interest of allowing the interpretability of machine learning models. In this work, we carry out an in-depth study of explainability analysis for the algorithms of the LAMDA (Learning Algorithm for Multivariate Data Analysis) family that have been used in the context of supervised and unsupervised learning. In particular, for the case of classification the LAMDA-HAD algorithm, and for the case of clustering the LAMDA-RD algorithm. For the explainability analysis, two classic methods from the explainability area were considered, LIME (Local Interpretable Model-Agnostic Explanation) and Feature Importance, and another one developed by us for the LAMDA family. In particular, our explainability method for LAMDA allows measuring the importance of each characteristic in a general way, and for each cluster. In general, the results obtained in both cases (classification and clustering) are satisfactory, especially because our explainability method for LAMDA gives an explainability similar to the traditional ones, but in addition, it can be given by cluster.

Cattle breeding has been one of the most important industrial sectors in the world, since it is related to food security and the survival of the human race. Cattle diagnostics is a fundamental procedure for cattle breeders because it allows them to make strategic decisions, such as timely treatment in case of any abnormality (e.g., weight gain in herds, in their paddocks). This article aims to present a system to diagnose weight loss or gain in cattle under a rotational grazing scheme, considering the health status of the animal and the pasture. The diagnostic system is based on a fuzzy classifier that uses fuzzy logic to define the rules that characterize the diagnostic process, and fuzzy reasoning to determine the current situation given an input. In addition, the fuzzy classifier optimizes the rules using genetic algorithms, which modify the membership functions, providing a more accurate system for diagnosis. We tested our proposal with experimental cases, with promising results. The accuracy metrics have high values, indicating a low error rate in terms of false positives. In general, the values of the quality metrics are very good, with an accuracy close to 100% and an Area Under the Curve close to 1.