Water　seepage　in　concrete　can　significantly　degrade　the　durability　of　hydraulic　concrete　structures.　Therefore,　this　paper　introduces　a　new　method　that　combines　the　percussion　method　with　deep　learning　techniques　to　detect　the　depth　of　water　seepage　in　concrete　structures.　Initially,　percussion　sound　signals　were　collected　for　different　water　seepage　depths.　Then,　the　proposed　one-dimensional　convolutional　bidirectional　gated　recurrent　unit　(BiGRU)　network　with　wide　first-layer　kernel　(1D-WCBGRU)　classifies　the　percussion　sound　signals　for　different　water　seepage　depths.　The　1D-WCBGRU　uses　a　wide　first　convolutional　kernel　to　extract　features　directly　from　the　original　percussion　signals　without　the　need　to　extract　features　manually.　Subsequently,　the　BiGRU　is　utilized　to　capture　long　short-term　information　from　the　data,　thereby　enhancing　feature　separability　and　improving　the　classification　accuracy　and　robustness　of　the　model.　Experiments　confirm　that　the　1D-WCBGRU　exhibits　excellent　performance　in　the　seepage　depth　detection　task　compared　to　traditional　learning　algorithms.