Data-Driven Risk Scoring for Dairy Farm Compliance Using Multimodal Sensor Streams and Federated Analytics

Abstract

Continuous compliance monitoring in dairy farming remains impeded by episodic inspection protocols and fragmented sensor data management. This paper proposes FedRS, a data-driven composite risk scoring framework that integrates multimodal sensor streams—body temperature, accelerometer-derived activity patterns, ambient ammonia concentration, and milk pH—through attention-weighted late fusion and gradient-boosted risk scoring to produce a continuous probability-of-violation signal at five-minute inference intervals. Risk scores are computed on-device at farm nodes and aggregated across heterogeneous farm populations through a privacy-preserving federated analytics protocol incorporating Graph Attention Network (GAT)-based cluster formation, DBSCAN outlier exclusion, and DP-SGD differential privacy. Evaluated on a 20-farm federated simulation derived from the Shahhet28121 livestock dataset, FedRS achieves 94.7% classification accuracy, 0.967 AUC-ROC, and a false-negative compliance violation rate of 1.9%, outperforming standard FedAvg by 3.4 percentage points while reducing per-round communication overhead by 97.7% to 4.25 KB. The framework maintains accuracy above 90% under sensor noise levels up to σ = 0.20, and ablation analysis confirms that GAT-based cluster-aware aggregation is the dominant performance driver. FedRS provides a technically rigorous, privacy-preserving, and communication-efficient foundation for automated dairy compliance monitoring systems compatible with low-bandwidth rural IoT deployments.

Keywords: Dairy compliance monitoring; Risk scoring; Federated learning; Multimodal sensor fusion; Graph attention network; Differential privacy; Internet of Things; Precision livestock farming