Abstract: Concrete structures are prone to damage under the coupled effects of long-term loading and environmental conditions, making early and accurate prediction of their failure status crucial for engineering safety. Traditional monitoring methods—relying on empirical thresholds or single-dimensional signal parameters—often suffer from low identification accuracy. To address this issue, this study conducts pressurized concrete tests to acquire synchronized load and acoustic emission (AE) full-waveform data. Mel-frequency cepstral coefficients (MFCCs) are extracted as feature parameters to construct a machine learning–based prediction model for failure status. The proposed method leverages the nonlinear time–frequency characteristics of AE signals to enhance representation of complex damage patterns. Experimental results demonstrate that the model effectively predicts the damage status of concrete; the average AUC is 0.9174. This approach provides a highly sensitive, intelligent diagnostic framework for structural health monitoring and real-time safety early warning in concrete structures.