Abstract
Inorganic chemistry laboratories act as core facilities for inorganic material synthesis, elemental analysis, and chemical reaction research. Nevertheless, the widespread use of reactive inorganic reagents (e.g., strong oxidants, reducing agents, heavy metal salts), corrosive substances, and high-temperature/pressure experimental processes often results in safety accidents including chemical reactions out of control, reagent splashes, toxic gas emissions, and equipment corrosion. Conventional monitoring systems, reliant on manual inspections and single-sensor alarms, suffer from detection delays, high false alarm rates, and inadequate emergency response. To address this, this paper proposes an intelligent safety monitoring and emergency decision-making system based on multimodal deep learning. It integrates visual imagery with multidimensional sensor data (temperature, humidity, toxic inorganic gas concentration, corrosive gas partial pressure, heavy metal ion concentration) to construct an enhanced YOLOv8-CBAM object detection model and a bidirectional long short-term memory (BiLSTM) temporal prediction model. Attention mechanisms enable multimodal data fusion, culminating in an emergency decision module designed through rule-based and case-based reasoning. Experimental results demonstrate that the enhanced YOLOv8-CBAM model achieves a 96.8%
[email protected] for detecting reaction flaring, corrosive splashing, toxic smoke, and non-compliant operations, representing a 3.2% improvement over the original YOLOv8. The BiLSTM model achieved a low MAE of 0.023 for sensor data prediction, outperforming traditional LSTM; post-multimodal fusion, safety state classification accuracy reached 98.2%, with the system's average emergency response time controlled within 12 seconds. This effectively enhances laboratory safety prevention and emergency response capabilities.
Download as PDF