The identification of mental tiredness which is based on EEG is basically limited by the problems of insufficient labeled data, large differences among different individuals, and high requirements for calculation ability. This research constructs a synergistic frame which uses an Improved Latent Diffusion Model (ILDM) to carry out frequency-perception diffusion in a compressed latent manifold, therefore it promotes spectral fidelity and at the same time greatly decreases calculation cost. For the purpose of guaranteeing effective state distinguishing, a light-weight 1D-UNet which contains multi-scale time convolutions and linear attention has been developed for actual putting into use. Under the validation of Leave-One-Subject-Out, the structure put forward by us obtains mean accuracy values of 76.60% on the SEED-IVG three-class task and 82.06% on the SADT two-class task, therefore it significantly surpasses the traditional GAN, VAE and standard DDPM models. The number-based analyses prove that the compounded signals keep the complicated function connection modes and rhythm vibrations which are necessary for strong tired degree measurement. In addition, the combination of Grad-CAM-based visual method connects deep learning forecast results with neurophysiology explanation ability, hence it proves that rear brain cortex areas have leading function in fatigue change processes. This whole combined design at the same time solves problem of data lacking and putting efficiency, therefore it pushes forward the dependability of automatic nerve physiology check systems.
