Machine learning techniques have become increasingly applied to non-destructive testing based on pulsed thermography. However, existing methods need to extract characteristic data manually. The present work addresses this issue by applying a bidirectional long-short term memory (Bi-LSTM) network to identify defects, predict defect depths, and reconstruct defective materials in three dimensions automatically based on raw cooling data sequences. The network is trained and tested based on data collected for stainless-steel specimens with multiple flat-bottom holes introduced at various specimen depths. A dual-task method and a single-task method were proposed based on Bi-LSTM network. A classification model and a regression model are constructed in the dual-task method for identifying defects and predicting defect depths. Only the regression network is implemented in a single-task method to more quickly obtain the same results based on a depth threshold. Both methods are demonstrated to achieve satisfactory accuracy in the 3D reconstruction of the defects in the testing specimen. In addition, higher pulse energy and faster acquisition frequency can promote the prediction accuracy. Then the results of Bi-LSTM were compared with the results of 1D CNN and MLP. To verify the generalization of the proposed method, CFRP specimen is employed for 3D reconstruction, which also performed with good results.