Software failure prediction is one of the most critical issues in software reliability research. And there are several crucial challenges for software failure prediction, which includes complex data, intrusive monitoring that brings potential risk to the system and monitoring costs. We proposed the FPbTI approach to predict software failure by using thermal images taken from outside of the hardware. In this approach, a data processing method is proposed and a convolutional neural network model is constructed. Additionally, corresponding experimental system was designed to verify the effectiveness of the approach. We experimentally demonstrate that our proposed approach achieves the goal of software failure prediction and shows remarkable performance.