An adaptive expert-in-the-loop algorithm for flock-specific anomaly detection in laying hen production

Laying hen egg production shows high day-to-day intra-flock and inter-flock variability due to environmental stressors and suboptimal welfare, which negatively impact egg production. While laying hen flocks maintain efficient egg production until 100 weeks of age, the detection of health and welfare issues becomes increasingly important to prevent long-term effects on production and consequently on farmer economics. The intrinsic non-stationarity of continuously streaming production data imposes challenges on anomaly detection, rendering the current solutions for anomaly detection unsuitable for flock- and farm-specific, adaptive, high-confidence anomaly detection. In this paper, we propose an adaptive expert-in-the-loop algorithm for early anomaly detection in daily laying hen egg production. The key point was to dynamically model flock-specific egg production curves, using incremental one-class support vector machines (OCSVM), and compare daily acquired production data to an expert-defined adaptive reference trajectory, while allowing incorporation of variables related to hen performance or environmental variables. Detected anomalies receive an anomaly score based on a predefined normalized score threshold. Expert feedback is asked in instances of low-confidence anomalies, to iteratively improve accuracy of the anomaly detection algorithm. The proposed model was trained and tested, using real flock and synthetic datasets. Incremental learning improved anomaly detection precision from 0.70 to 0.81 compared to the initial OCSVM model. Expert feedback further refined the balance between sensitivity and precision, with an F1-score of 0.93 with 13% of expert feedback, thereby lowering false alarm ratios, while improving anomaly detection capabilities. Although this algorithm focusses on egg production, it can be adapted to detect anomalies in other production features, such as egg weight.