Abstract: ‍ ‍Diseases related to the human respiratory system are often accompanied by abnormalities in the respiratory depth and rhythm. Respiratory signal monitoring and respiratory pattern recognition are very important in the medical and health fields， especially for applications such as sleep monitoring and disease prediction. Among the various methods utilized， non-contact impulse radio ultra-wideband （IR-UWB） radar is gradually becoming one of the most critical sensing technologies in the field of sleep health monitoring because of its excellent range resolution and penetration ability， as well as the advantages of all-weather， all-time， safe， and non-invasive detection. However， a complex indoor measurement environment poses limitations and challenges to the accurate extraction of respiratory pattern features. Traditional radar respiratory pattern recognition algorithms mainly focus on one-dimensional respiratory time and frequency domain features， whereas the echo information targeted by IR-UWB radar is scattered within multiple range bins， which causes a low accuracy when using one-dimensional features for recognition. Therefore， this paper proposes a multi-domain feature fusion respiratory pattern recognition method with time-distance information for IR-UWB radar， which focuses on the characteristics of a signal model in which human respiration is an extended target in distance and slowly fluctuates in time. Based on extracting the time and frequency domain features of one-dimensional respiratory signals， the algorithm was used to further explore the potential respiratory pattern morphological features in two-dimensional respiratory radar images. The non-contact detection and recognition of respiratory patterns were achieved based on the multi-domain feature information of respiratory signals in time and distance. An image-processing method was proposed to detect respiratory time-distance bands in radar images， Phase-matrix image processing was used to obtain image features and address the problem of local adhesion in images affected by abnormal respiratory rhythms， which make it difficult to extract respiratory cycles. Finally， in an experiment involving the machine learning classification and recognition of six respiratory patterns， the multi-domain features extracted by this algorithm achieved a recognition accuracy of 96.3%. The experimental results showed that this method could capture richer spatiotemporal information about respiratory patterns， providing a new approach for respiratory pattern recognition. It demonstrated better respiratory-pattern recognition performances than traditional methods， indicating that the proposed approach has potential for sleep-state monitoring and assisted disease diagnosis.