Integrating Heterogeneous Sensor Streams with Fusion-Enhanced Deep Learning for Robust Anomaly Detection in Tobacco Preservation Systems

This study addresses the challenges of heterogeneous multi-source sensor data, complex temporal patterns, diverse anomaly types, and severe noise interference in tobacco storage environments through in-depth modeling. A Fusion-Enhanced Deep Network (FEDN) is proposed and applied to achieve highly reliable anomaly identification in tobacco storage systems. First, considering the differences in sampling frequency, physical dimensions, and noise distribution among various regions and types of environmental sensors, a unified spatiotemporal modeling framework for multi-source heterogeneous sensing is constructed. Simultaneously, a standardized fusion expression of multiple parameters such as temperature and humidity is achieved through exponential moving average and physical coupling functions. Second, considering the multi-timescale characteristics of tobacco storage environments, including short-period perturbations, diurnal cycles, and long-term heat accumulation, a multi-scale temporal convolution-attention fusion structure is designed. This structure enables collaborative modeling of features across temporal resolutions. Spatially, a graph-based encoder is introduced to characterize the topological dependencies between sensor nodes. Furthermore, stability constraints are used to improve the numerical stability of spatial feature aggregation. At the crossmodal modeling level, a cross-modal attention mechanism is constructed between temperature, humidity, airflow, and smoke concentration. In a real-world industrial setting, a 180-day continuous data acquisition experiment was conducted on an automated warehouse of Zhejiang Tobacco. Ninety-six environmental sensor nodes were deployed at the data acquisition level, accumulating approximately$4.2 \times 10^{8}$high-density timeseries data points. These data addressed six typical problematic operating conditions: localized heat accumulation, abnormally low humidity, severe fluctuations, early indicators of fire, sensor malfunction, and irregular airflow. Models including LOF, Isolation Forest, LSTM-AE, TCN-AE, GCN-AE, CNN-LSTM, and Transformer-AE were used in comparative experiments. The proposed approach was rigorously evaluated using several metrics including Precision, Recall, F1-score, AUC, and FAR. According to experimental results, FEDN received an overall test AUC of 0.978 and an F1-score of 93.7 %. This result represents an improvement of approximately 3.5 percentage points compared to the suboptimal Transformer-AE model, and maintains stable detection performance above 90 % across three types of time-series anomalies: short-period, medium-period, and long-period. In anomaly type identification, the accuracy rate for recognizing early fire signs reaches 98.0 %. In strong noise robustness testing, when the Gaussian noise intensity increases to$\sigma=0.20$, the model performance degradation is only 4.7 %.