The non-invasive blood glucose detection technique is an indirect method of measuring glucose levels in the blood, which is safe, fast, and non-invasive without damaging human tissues, breaking the limitations of traditional blood glucose detection, and has important research value. The photoplethysmography signal, containing various physiological and pathological information, is widely used in various clinical studies and is also the focus of attention in the current implementation of the non-invasive glucose detection technique. Current studies of non-invasive blood glucose detection based on photoplethysmography signals have only considered the contribution to system modelling when the time or frequency domains act alone. Although the time domain analysis of the signal can describe the variation of the PPG signal amplitude with time, it cannot visually reflect the energy distribution of the PPG signal frequency. Therefore, the signal analysis of a single domain cannot fully express the PPG signal, which leads to information loss. When using frequency domain analysis to extract the signal spectrum, it is necessary to use all the time domain information of the signal, which is a global transformation and may result in the loss of signal characteristics at a specific time or in a specific frequency band. In summary, this paper proposes a new method for non-invasive blood glucose detection based on the integrated time-frequency domain analysis of photoplethysmography (PPG), using a parallel time-frequency domain method to consider the association between the photoplethysmography signal and blood glucose, a cluster analysis method is used to extract representative waveforms in the time domain of the PPG signal, analyze the correlation between the waveform features and blood glucose, and determine the time domain feature parameters of the waveform. On this basis, the pulse waveform time domain signal is converted to the frequency domain using the Fast Fourier Transform, and the spectral information is studied using principal component analysis to establish the frequency domain characteristic quantities. The BP neural network-based non-invasive blood glucose detection model is constructed by extracting the time-frequency domain feature parameters from the waveform signals obtained through the Oral Glucose Tolerance Test (OGTT) and using the invasive blood glucose concentration detected in real time as a reference. At the same time, in order to improve the accuracy of the model and achieve model optimization, a genetic algorithm is applied to the model for the second correction, and the final MAE and RMSE of the test set reach 1.13 and 1.42 mmol·L-1. The results of Parkers CEG show that the prediction results in the A and B regions accounted for 80.3% and 19.7%, respectively, which indicates that the method has good prediction accuracy and provides a theoretical basis for the feasibility of daily non-invasive blood glucose monitoring. It is beneficial to improve the detection and monitoring systems for diabetes, to judge the condition better comprehensively, and to prevent, guide, and treat diabetes promptly.