Multi-algorithmic approach for detecting outliers in cattle intake data

Abstract

Monitoring cattle feed intake is crucial for evaluating animal health, productivity, and farm profitability. In particular, an abnormal intake is related to the cattle activity. Therefore, outlier detection forms the basis for intake monitoring. This study employed multiple algorithms ranging from statistics to deep learning to detect outliers in time-series data of cattle intake. We used five models implementing mean + standard deviation, moving average, box plot, time series decomposition, and autoencoder, and attempted to enhance the detection performance by a voting system to combine more than one model. Both box plot and time-series decomposition demonstrated high accuracy (over 95 %) and F1-score (harmonic mean of precision and recall). Thus, it reliably distinguished normal values from outliers. Moving average exhibited a high true-skill statistic (TSS), thereby rendering it suitable for outlier detection. The voting system gave F1 and TSS scores of 0.49 and 0.65, respectively. Thus, it enhanced the detection performance compared with the individual model. These results demonstrate that the performance metrics vary depending on the type of algorithm. This, in turn, highlights the need to select algorithms adapted to the monitoring objectives. The algorithmic selection can be complemented by a voting system. This demonstrates its potential for generating a reliable database with accurate outlier detection and aiding decision-making by livestock producers.

Keywords

• Anomaly detection
• Cattle intake
• Machine learning
• Statistics
• Voting process