The aging of nuclear reactors presents a substantial challenge within the field of nuclear energy. Consequently, there is a critical demand for field reconstruction techniques capable of obtaining comprehensive spatial data about the condition of nuclear reactors, even when provided with limited observer data. It is worth noting that prior research has often neglected to account for the impact of noise and changes in sensor states that can occur during actual production scenarios. In this paper, the so called Noise and Vibration Tolerant ResNet (NVT-ResNet) is proposed to tackle these challenges. By introducing noise and vibrations into the training data, NVT-ResNet is able to learn the tolerance thus exhibits robustness for the field reconstruction. The influence of sensor numbers on the model’s performance is also investigated. Numerical results convincingly demonstrate that even with limited sparse sensors exposed to a noise with magnitude of 5% and vibrations, NVT-ResNet consistently achieves a reconstruction field of relative $L_2$ error within 1% and relative $L_∞$ error of less than 5% in average sense. Additionally, NVT-ResNet exhibits remarkable computational efficiency, with field reconstruction taking only microseconds. This makes it a viable candidate for integration into online monitoring systems, thereby enhancing the safety performance of nuclear reactors.