To provide a robust working environment for TENGs, most TENGs are designed as sealed structures that isolate TENGs from the external environment, and thus their operating conditions cannot be directly monitored. Here, for the first time, we propose an artificial neural network for interface defect detection and identification of triboelectric nanogenerators via training voltage waveforms. First, interface defects of TENGs are classified and their causes are discussed in detail. Then we build a lightweight artificial neural network model which shows high sensitivity to voltage waveforms and low time complexity. The model takes 2.1 s for training one epoch, and the recognition rate of defect detection is 98.9% after 100 epochs. Meanwhile, the model successfully demonstrates the learning ability for low-resolution samples (100 × 75 pixels), which can identify six types of TENG defects, such as edge fracture, adhesion, and abnormal vibration, with a high recognition rate of 93.6%. This work provides a new strategy for the fault diagnosis and intelligent application of TENGs.