As a significant carrier of quantum information, the studies of qubits are providing more opportunities in emerging research fields such as quantum computation, quantum simulation and quantum metrology. However, due to the existence of a noise environment, such as magnetic signal fluctuation, the resulted decoherence of qubits limits its capabilities in the above-mentioned fields. Acknowledging the noise information can help us to break the limitation and improvethe applications of qubits, based on which crucial optimization of the customized dynamical decoupling protocols and suppression of the noise could be well realized. Thus, an accurate and efficient approach for analyzing the spectral information of the noise environment is required. Because of the difficulty of function inverse solution, conventional analytical methods based on approximation techniques can not accurately resolve the noise spectra from the interrogation-time-domain measurements of qubits.Recently, significant advance has been made in deep learning, which has been widely used for quantum information processing. In this paper, we propose a deep-learning-based method for noise spectral analysis of qubits. This method only needs to input thedecoherence curves into the deep-learning model to predict the existed environmental noise spectra. Through a series of iterative studies, this method can extract the potential mapping relation between the decoherence curves of qubits and corresponding noise spectra. We numerically simulate the deep-learning-based analysis process and demonstrate the good performance of our method. Moreover, benefited from the deep-learning-based algorithm, compared with the conventional method, the accuracy and efficiency of our method are much better. Our method also provides a new technique for other noise spectral analysis tasks and could be easily applied to a wide range of quantum systems.