Traditional near-infrared spectrometers are bulky and expensive， limiting their widespread application. Based on Hamamatsu's MEMS-FPI near-infrared spectroscopy sensor C14272， this research successfully developed a miniaturized near-infrared spectrometer operating in the 1350-1650 nm wavelength range. The system adopts a modular architecture， integrating efficient signal acquisition and data processing units， significantly reducing system costs. Through systematic performance evaluation， including wavelength accuracy， spectral resolution， dynamic range， and other metrics， combined with a temperature characteristic compensation algorithm that effectively suppresses temperature drift effects， experimental results demonstrate that this spectroscopic analysis system possesses excellent stability and reproducibility， with a spectral resolution of approximately 15.2 nm. Combined with a convolutional neural network algorithm， the glucose solution concentration prediction model exhibits high accuracy （R²>0.99） and low error rates （RMSE<0.2%） when detecting unknown samples. The developed miniaturized near-infrared spectrometer provides an economical and efficient solution for biomedical detection and food quality analysis.