An initiating event (IE) is an event that may lead to core damage in a nuclear power plant (NPP), and being able to identify an IE is crucial in determining what actions to take. This process requires examining the data from the widespread NPP sensors. Any fault in these sensors can be a severe issue that may defer the mitigation action that would bring the NPP back into safe shutdown conditions. In this study, we focus on addressing two issues in this regard: (1) event identification with sensor faults, and (2) faulty sensor signal reconstruction under an IE with the help of two deep learning (DL)-based schemes. Through the superior capability of DL in feature extraction, discriminant information can be extracted to enable valid event identification even when the faulty sensors have been discarded. Owing to the nonlinear nature of the DL network model, the process information for sensor reading generation and the interrelations among the sensors in the event recordings can be used to reconstruct faulty sensor readings without the need for deterministic trend removal. Results from detailed experiments using data generated by a simulator of the Maanshan NPP in Taiwan are provided to illustrate the efficacy of the proposed approaches.