Junction temperature monitoring (TJM) is critical in the reliability assessment of insulated-gate bipolar transistor (IGBT) modules. However, existing TJM methods are usually disturbed by aging. To address this, this paper proposes a novel approach for TJM, which utilizes convolutional neural network (CNN) algorithm and the turn-off collector current I_C. Firstly, the dynamic characteristic model of turn-off I_C and the thermal characteristics are analyzed. To examine the correlation between the turn-off I_C and the junction temperature T_j, a mathematical model of the I_C is established. Through the implementation of the double-pulse tests, the connection between the turn-off I_C and T_j is verified, and other factors related to turn-off I_C are also analyzed. Finally, a TJM method is proposed based on convolutional neural network (CNN), which can decouple the effects of aging.