To address the issues of low signal-to-noise ratio, high background noise, and non-stationarity in detecting concrete structure source signals based on piezoelectric wave method, a piezoelectric signal filtering method based on K-singular value decomposition (K-SVD) to update the dictionary was proposed, and the damage of concrete structures were identified. Firstly, piezoelectric signals from concrete structures in both cracked and intact states were collected and classified. Secondly, the acquired piezoelectric signals were filtered, and the the results using the K-SVD dictionary learning method were compared and analyzed with the unfiltered results to evaluate the applicability of the K-SVD dictionary learning filtering method. Finally, the filtered piezoelectric signals using ResNet were classified and recognized. The results show that the method based on K-SVD dictionary learning and ResNet can stably identify the piezoelectric signals of internal damage in concrete structures. The accuracy of damage signal recognition in training set and test set is 93.25% and 92.38%, respectively. The recognition accuracy of lossless signal is 95.41% and 94.67%,respectively, which is more than 10 percentage points higher than that of unfiltered signal. The effective damage identification in concrete bridge structures through the integration of K-SVD dictionary learning and ResNet has achieved the localization of internal damage areas in the concrete structures. The research findings present a novel approach to data processing in the health monitoring of concrete bridge structures.