In　this　letter,　a　deep-learning-assisted　signal　detector　is　developed　for　communication　systems　with　unknown　channel　models.　By　embedding　domain　knowledge　into　a　self-attention　model,　a　novel　detection　unit　is　devised　that　enables　both　reliable　estimation　and　fast　training.　Furthermore,　a　sliding-window　structure　is　in　combined　use　with　the　detection　unit　to　realize　real-time　signal　recovery.　We　evaluate　the　performance　of　the　proposed　detector　using　a　chemical　communication　experimental　platform,　and　show　the　superiority　of　our　design　in　terms　of　detection　accuracy　as　well　as　implementation　complexity.