Top

Edge-Aware KAN-GraphMixer Forecasting Model for Abnormality Identification in Intelligent Logistics Cold Boxes

Authors

Files

PDF

Abstract

Abnormality identification in intelligent logistics cold boxes requires reliable interpretation of coupled sensor signals under transient handling, uneven thermal response, and intermittent communication. This study proposes an edge-aware KAN- GraphMixer forecasting model that combines nonlinear channel transformations with lightweight sensor-graph mixing. A Kolmogorov-Arnold network represents channel-specific response functions, while graph mixing propagates information among thermally related sensing locations. Multi-step forecasts are converted into node-level residuals, and abnormality decisions incorporate residual magnitude, persistence, and observation confidence. An illustrative evaluation configuration comprising 42,600 observation windows covers normal transport, short and prolonged door exposure, delayed pre-cooling, refrigeration interruption, and sensor drift. The reported configuration achieves a 15-min temperature forecasting mean absolute error of 0.286 °C and a root mean square error of 0.438 °C. Abnormality identification reaches 97.5% precision, 96.1% recall, and a 96.8% F1-score. Compared with the graph-mixing baseline, forecasting error decreases by approximately 10.1%, and F1-score increases by 2.4 percentage points. At a false alarm rate of 1.9%, the median warning lead time is 11.6 min. Local inference requires 8.7 ms per window, with a quantized parameter footprint of 0.84 MB. Component comparisons indicate that both nonlinear transformation and graph interaction contribute to performance, while packet loss and shared calibration reduce identification accuracy. The numerical findings describe a proof-of-concept configuration whose practical effectiveness requires independently collected operational records, reproducible training settings, and gateway profiling. They indicate potential for compact predictive monitoring within the stated sensing and calibration conditions.

PDF

Details