Multi-sensor fire detection technology has been widely adopted in practical applications; however, existing methods still suffer from high false alarm rates and inadequate adaptability in complex environments due to their limited capacity to capture deep time-series dependencies in sensor data. To enhance robustness and accuracy, this paper proposes a novel model named BiLSTM-LN-SA, which integrates a Bidirectional Long Short-Term Memory (BiLSTM) network with Layer Normalization (LN) and a Self-Attention (SA) mechanism. The BiLSTM module extracts intricate time-series features and long-term dependencies. The incorporation of Layer Normalization mitigates feature distribution shifts across different environments, thereby improving the model’s adaptability to cross-scenario data and its generalization capability. Simultaneously, the Self-Attention mechanism dynamically recalibrates the importance of features at different time steps, adaptively enhancing fire-critical information and enabling deeper, process-aware feature fusion. Extensive evaluation on a real-world dataset demonstrates the superiority of the BiLSTM-LN-SA model, which achieves a test accuracy of 98.38%, an F1-score of 0.98, and an AUC of 0.99, significantly outperforming existing methods including EIF-LSTM, rTPNN, and MLP. Notably, the model also maintains low false positive and false negative rates of 1.50% and 1.85%, respectively. Ablation studies further elucidate the complementary roles of each component: the self-attention mechanism is pivotal for dynamic feature weighting, while layer normalization is key to stabilizing the learning process. This validated design confirms the model’s strong generalization capability and practical reliability across varied environmental scenarios.

Multi-sensor fire detection technology has been widely adopted in practical applications; however, existing methods still suffer from high false alarm rates and inadequate adaptability in complex environments due to their limited capacity to capture deep time-series dependencies in sensor data. To enhance the robustness and accuracy of fire detection, this paper proposes a fire detection model based on a Bidirectional Long Short-Term Memory network with Layer Normalization and Self-Attention (BiLSTM-LN-SA). The model employs a Bidirectional LSTM (BiLSTM) to autonomously extract intricate time-series features and long-term dependencies from multi-sensor data. Furthermore, Layer Normalization(LN) is introduced to effectively mitigate feature distribution shifts across different environments, thereby improving the model's adaptability to cross-scenario data distributions and generalization capability. Coupled with a self-attention mechanism that dynamically evaluates the importance of features at different time steps, the model adaptively enhances fire-critical information and achieves deeper dynamic process-aware feature fusion. Experimental results on a real-world fire dataset demonstrate that the BiLSTM-LN-SA model effectively identifies fire events, exhibiting superior detection performance.

Dust concentration during intermittent processing of powder and granular materials is characterized by high-frequency abrupt changes, local accumulation, and complex coupling among multiple sensors. Existing forecasting models still exhibit limitations in modeling global dependencies and characterizing local trends. To address these issues, this paper proposes an iTransformer-based dust concentration forecasting model that integrates a dual-stage feed-forward network and a DLinear branch. With iTransformer as the backbone network, the proposed model captures the coupling relationships among multi-source sensing signals through variate-wise modeling. A progressive dual-stage feed-forward feature refinement mechanism is constructed to enhance the model’s representation capability for transient variations and peak fluctuations in dust concentration. In addition, a collaborative modeling framework consisting of an iTransformer main branch and a DLinear auxiliary branch is designed to jointly learn global nonlinear features and local linear trends. An adaptive gated fusion mechanism is further introduced to dynamically allocate the contribution weights of different branches according to sequential characteristics. Experiments were conducted on a public 1 Hz smoke-sensing dataset, which was used as a proxy benchmark for high-frequency multivariate PM2.5 forecasting rather than direct industrial dust data. Under the setting of a 300-step input length and a 60-step forecasting horizon, the proposed model achieves an MSE of 1.8292 × 10⁻³, an MAE of 0.0334, an RMSE of 0.0428, an MAPE of 0.0177, and an R² of 0.9744, outperforming the compared baseline models in overall performance. The results indicate that the proposed method improves overall forecasting accuracy and provides a methodological reference for sensor-driven particulate concentration forecasting and early warning, while further validation using field data from actual powder and granular material processing workshops is still required before practical deployment.