A Time-Series AI Framework for Predictive Risk Assessment from Heterogeneous Sensor Data
This paper presents a time-series AI framework for predictive risk assessment using heterogeneous, irregularly sampled sensor data. The framework targets monitoring scenarios in which environmental covariates are available at high frequency, while the target variable is sparse, delayed, or obtainable only through laboratory analysis. The proposed pipeline integrates historical environmental data, IoT sensor streams, feature engineering, temporal windowing, domain alignment, and sequence-learning models within a unified forecasting process. The framework is evaluated in a precision livestock case study to predict the Total Bacterial Count in buffalo milk, a proxy indicator of microbiological risk. Sparse real TBC measurements are combined with Copernicus reanalysis data, local environmental sources, and farm IoT sensors. Recurrent and Transformer-based models are trained on sliding temporal windows and evaluated on both real and simulated TBC targets. Results show that datasource quality, feature representation, and look-back window length strongly affect predictive performance. Copernicus-based data and moderate temporal windows provide robust results, while the sparsity of real microbiological observations remains the main limiting factor. The proposed framework supports the transition from retrospective model training to sensor-based operational inference for early risk assessment.