Abstract: ‍ ‍Blood pressure is a critical physiological indicator for assessing cardiovascular health， and regular monitoring aids early diagnosis and intervention of cardiovascular diseases. Compared to traditional electronic sphygmomanometer measurement methods， millimeter wave radar provides non-contact measurement advantages， showing promising future applications. This study presents a neural network-based method using a millimeter-wave radar for accurate blood pressure measurement and waveform reconstruction in a non-contact manner. Initially， radar signals were preprocessed using an extended differentiate and cross-multiply algorithm and average filtering to effectively extract phase information from the radar echoes and eliminate constant phase components. Wavelet filtering was then employed to eliminate high-frequency noise and baseline drift from the signals， obtaining high-quality pulse wave signals. Subsequently， a two-stage progressive feature fusion and mapping network with an encoder-decoder structure was constructed to establish the mapping between the pulse wave characteristics and blood pressure， enabling accurate blood pressure measurement and waveform reconstruction. In the first stage， MultiResUNet was used as the backbone network to extract and fuse multi-scale features of the pulse wave while introducing a self-attention mechanism between multi-resolution blocks to explore long-distance dependencies between feature vectors， thus accurately reconstructing the blood pressure waveform. In the second stage， the model automatically extracted deep features of the pulse wave using the encoder trained in the first stage. The features were then further integrated and mapped using convolutional neural networks and long short-term memory networks to estimate systolic and diastolic blood pressures. Finally， the proposed method was validated using the radar vital signs dataset， yielding a measurement error of 3.49$\pm$5.75 mmHg for systolic blood pressure and 2.40$\pm$3.59 mmHg for diastolic blood pressure， meeting the A-grade requirements of the British Hypertension Society standards. Additionally， the reconstruction error of the blood pressure waveform was 3.33 mmHg with a deviation rate of 3.74%， further demonstrating the effectiveness of this method in blood pressure waveform reconstruction.