Abstract: The traditional method for detecting generator stator slot wedge tightness relies on manual tapping and subjective judgment, which suffers from heavy dependence on human experience and is prone to causing operator fatigue. To address this limitation, this paper proposes a fault detection method for generator stator slot wedge tightness based on multidimensional joint TEO-MFCC features and a hybrid CNN-LSTM model. First, sound signals corresponding to different wedge states are acquired by tapping the slot wedge using a self-developed electromagnetic excitation device. Second, the MFCC extraction process is enhanced by incorporating the Teager Energy Operator (TEO), and three acoustic feature sets—MFCC, delta-MFCC (DT-MFCC), and delta-delta-MFCC (DDT-MFCC)—are constructed to capture dynamic energy characteristics. These features are then concatenated to form the multidimensional joint TEO-MFCC representation, whose discriminative capability is validated via inter-class visualization. Third, the joint TEO-MFCC features are fed into a hybrid CNN-LSTM network: convolutional layers extract local temporal-spectral patterns, while the LSTM layer models long-range temporal dependencies to perform end-to-end fault diagnosis of wedge tightness. Experimental results demonstrate that the proposed method achieves an accuracy of 98.7% in diagnosing stator slot wedge tightness—improving upon standalone CNN and LSTM baselines by 8.6% and 4.9%, respectively—thereby validating its effectiveness under complex operational conditions. This work provides a practical and reliable reference for intelligent fault detection of generator stator slot wedge tightness.