Fast Re-Trainable Attention Autoencoder for Liquid Sensor Anomaly Detection at the Edge

Modern life - science and chemistry laboratories handle highly reactive liquids such as strong acids and bases, organic solvents, and powerful oxidisers. Small deviations in temperature, concentration, stirring speed, or dissolved - oxygen level can trigger unpredictable behaviour that release s toxic gases, generates intense heat, or causes explosions. These events place personnel, facilities, and property at serious risk. Statistics from the U.S. Chemical Safety Board, covering 2013 to 2023, show that liquid - chemical leaks make up about thirty percent of all laboratory incidents; forty - two percent of those incidents lead to human exposure, and twelve percent require building evacuation. Each liquid has its own distribution of normal physicochemical values, so baseline sensor readings change from one experiment to another. Redesigning and relabelling a multi - class model for every new setup is impractical. Current monitoring still relies on visual checks and single - sensor alarms, which do not capture correlations among sensors. Cloud - based IoT solutions are often blocked in high - security laboratories because data must remain on site and Internet latency cannot be guaranteed. An edge - resident intelligent system that processes multimodal data in real time and issues early warnings inside the laboratory network is therefore required.