It is ready to cause the interface delamination, macroscopic fracture and spalling failure of thermal barrier coatings (TBCs) due to their complexity of architecture and harsh service environment. In this paper, the failure process of TBCs under three-point bending (3PB) load was monitored in real time via acoustic emission (AE) technology, and the damage failure modes of TBCs were identified based on micro-morphology analysis of AE parameters and K-means cluster. The waveform characteristics of four failure modes were analyzed by fast Fourier transform and wavelet packet transform. The macroscopic fracture or spalling failure signals have no obvious frequency band, while the corresponding frequency components of substrate deformation, surface vertical crack, sliding interface crack and opening interface crack are 62.5?125.0 kHz, 187.5?250.0 kHz, 250.0?312.5 kHz and 375.0?437.5 kHz, respectively. The method of deep machine learning was used to process the in-situ acoustic emission signals. The wavelet energy coefficient was extracted as a characteristic vector of the Back propagation neural network, and the advantages and disadvantages of the model were evaluated by convergence curve, confusion matrix, Receiver operating characteristic curve and F1 value, thus realizing the discrimination of failure modes of TBCs under 3PB test and providing a reference value for failure prediction and life assessment of thermal barrier coatings.