Water　in　concrete　significantly　affects　its　durability,　so　detection　of　water　content　in　concrete　is　essential　to　ensure　its　durability　and　safety.　This　paper　introduces　a　method　for　detecting　moisture　content　in　concrete　structures　utilizing　percussion　and　deep　learning　techniques.　The　method　deploys　a　deep　neural　network　that　automatically　classifies　moisture　content.　A　two-stream　convolutional　bi-directional　long　short-term　memory　network　(TS-CBLSTM)　directly　processes　the　acquired　percussion　acoustic　signals　with　different　moisture　content.　The　TS-CBLSTM　employs　a　two-stream　convolutional　operation　to　extract　features　inherent　in　the　two　channels　of　the　original　audio.　Subsequently,　a　bi-directional　long　short-term　memory　(BiLSTM)　block　captures　the　connectivity　of　intrinsic　features,　thereby　enhancing　feature　separability.　This　approach　improves　the　classification　accuracy　and　robustness.　The　experimental　results　show　that　TS-CBLSTM　performs　brilliantly　in　concrete　moisture　content　detection　with　100%　classification　accuracy.　Furthermore,　the　intensive　study　of　TSCBLSTM＇s　noise　immunity　and　adaptability　confirms　that　it　outperforms　conventional　algorithms.