Neural network shows certain superiority in the field of dynamic light scattering inversion. However, the robustness and accuracy of the inversion results can be influenced by different training samples and network models. To solve this problem, an inversion method combining information bias weighting (IBW), convolutional neural network (CNN), and Transformer encoder (CNN-Transformer) is proposed. This method establishes a three-layer framework consisting of training sample generation, local feature extraction, and global dependency modeling. It uses the particle size distribution (PSD) obtained from IBW inversion as prior information to generate reliable training samples and utilizes the CNN-Transformer model to extract local features of the electric field autocorrelation function and perform global dependency modeling, thereby achieving an accurate PSD. The inversion results of simulated data with four different widths indicate that, compared with Tikhonov-PGRNN and IBW, the CNN-Transformer method achieves higher accuracy, stronger noise resistance, and better bimodal resolution for different particle systems. The experimental data inversion further validates the conclusions drawn from the simulated data. The peak position errors of 100, 203, 555 nm unimodal and 33 nm/203 nm, 300 nm/502 nm bimodal particle systems inverted by CNN-Transformer are 0, 1.48%, 0.90%, and 2.42%/0.06%, 1.27%/0.86% respectively, which are better than the inversion results of Tikhonov-PGRNN and IBW.