Pulse diagnosis， as a core diagnostic method in traditional Chinese medicine， has long faced challenges such as strong subjectivity and difficulty in standardization. To address these issues， this paper proposes a pulse diagnosis sensing system based on a tri-channel U-shaped Micro-Nano Fiber （MNF） sensor. The system can synchronously acquire pulse signals from the cun， guan， and chi positions at different pressing depths （floating， middle， and sinking）， comprehensively capture subtle variations in pulse waveforms， and integrate deep learning models to achieve accurate pulse recognition， thereby enhancing the objectivity and digitization of traditional Chinese medicine diagnosis.

To construct a synchronous signal acquisition system， three well matched U-shaped MNFs were embedded in a Polydimethylsiloxane （PDMS） film and fixed on a Polycarbonate （PC） substrate to form sensing units， which were arranged on the cun， guan， and chi parts respectively. MNF was prepared using flame heating cone method， with a diameter of about 3 μm and a bending radius of 2 mm. Its structural parameters were optimized through COMSOL simulation to achieve a balance between evanescent field effect and mechanical strength， ensuring both adaptation to radial artery pulsation and stable and accurate detection. The PDMS package thickness was optimized through ANSYS finite element simulation and ultimately determined to be 100 μm to achieve the best balance between mechanical strength and sensitivity. Three sensing units are integrated into a customized 3D printed silicone platform， which has good biocompatibility and can adapt to individual physiological differences. In the experimental section， a tri-channel Photoelectric Detection （PD） module was used to convert the optical signals into electrical signals， and the dynamic waveforms were displayed in real time on an oscilloscope. To simulate the three pulse conditions of “floating， middle， and sinking” in traditional Chinese medicine， an airbag cuff loading system was designed to apply different levels of pressure to the pulse positions， thereby enabling signal acquisition. Through signal processing， multiple physiological parameters including Rise Wave Transit Time （RWTT）， Upstroke Time （UT）， Pulse Transit Time （PTT）， Left Ventricular Ejection Time （LVET）， etc. are extracted. In terms of intelligent recognition， a dataset containing 1 260 annotated pulse wave data was constructed， covering seven typical pulse types： normal pulse， forceful pulse， rapid pulse， flooding pulse， slippery pulse， tense pulse， and thready pulse. A deep learning model combining one-dimensional Convolutional Neural Network （1D-CNN） and Bidirectional Gated Recurrent Unit （BiGRU） is proposed， where CNN is used for local feature extraction and BiGRU is used to capture temporal dependencies， thereby achieving efficient classification of multi-channel pulse signals.

Mechanical performance tests show that the sensitivity of the three channel MNF sensor in the pressure range of 0~7 kPa is 11.49% kPa^-1、11.64% kPa^-1 and 11.13% kPa^-1， respectively. The response time and recovery time are 22.7 ms/9.4 ms、31.2 ms/10.6 ms and 29.8 ms/8.5 ms， demonstrated favorable sensitivity and a high level of consistency. After 600 cycles of loading unloading， the signal remained stable， verifying the reliability of the packaging structure. Compared with existing flexible pulse sensors， this device has advantages in sensitivity， channel count， and response speed.In the pulse diagnosis experiment， the system can simultaneously obtain pulse signals from the cun， guan， and chi parts， and clearly characterize the waveform differences at different compression depths. As the pressure increases from shallow to deep， the pulse wave gradually flattens， UT shortens， and the rate of rise accelerates， indicating physiological changes in reduced vascular wall compliance. Through waveform analysis， objective parameters of seven common pulse patterns were successfully extracted， which were highly consistent with clinical manifestations in volunteer experiments. For example， diarrhea patients showed thready pulse， while cold patients showed tense pulse， demonstrating the effectiveness of the system in capturing pulse patterns. In the intelligent recognition experiment， the deep learning model achieved 91.92% and 92.01% accuracy in classifying left and right hand pulse signals， respectively， when only using intermediate data； When integrating data from floating， medium， and sinking phases， the accuracy significantly improves to 96.61% （left hand） and 96.07% （right hand）. Compared with existing classification methods based on manual features or single networks， this method exhibits higher classification performance on similar data scales， demonstrating the superiority of combining 1D-CNN and BiGRU in complex pulse recognition.

The flexible pulse diagnosis and perception system based on three channel U-shaped MNF proposed in this study can achieve synchronous acquisition of pulse waves from multiple parts and depths of compression， and combine deep learning to achieve high-precision pulse recognition. This method effectively improves the objectivity and standardization level of traditional Chinese medicine pulse diagnosis， and has broad application prospects in remote medical care， digital diagnosis and treatment， and chronic disease management.