Research on anomaly detection model of warehouse sensor data based on hybrid Fourier-temporal fusion network

Detecting anomalies in warehouse sensor data is critical for improving intelligent logistics management efficiency and ensuring the safety of stored goods. To overcome the limitations of traditional transformers in capturing time-domain dynamics and mining periodic features in multidimensional heterogeneous sensor data, we propose a Hybrid Fourier-Temporal Fusion Network (HyFT-Net). Built upon the transformer architecture, HyFT-Net integrates the dynamic sliding window for enhanced local temporal modeling of non-stationary signals and a fast Fourier transform frequency-domain channel attention mechanism to extract periodic spectral features more effectively. By fusing spatial-temporal dynamics with frequency-domain features via a dual-path collaboration mechanism, the model achieves superior abnormality detection performance, particularly on temperature and humidity data, compared to long short-term memory, transformer, and SparseTSF models. Specifically, for temperature, it reduces mean squared error by 10.00%–33.33% and mean absolute error by 10.37%–15.38%, while precision increases by 1.07%–4.67% and F1 increases by 0.85%–4.42%, demonstrating robust multi-sensor anomaly detection capabilities.