With the widespread application of deep learning in time series analysis, the predictive performance of models has been significantly improved. However, the “black-box” nature of these models still limits their trustworthiness and transparency in practical applications. Although many interpretability methods offer insights into model behavior, considerable limitations remain when handling complex time series data, particularly in tasks involving high-frequency components or long-period fluctuations. To address this challenge, this study proposes a time series interpretability method that integrates frequency-domain perturbation with time series segmentation. By employing short-time Fourier transform and dynamic time warping algorithms, the study first segments the time series, and then generates frequency-domain perturbations using generative adversarial network (GAN) to analyze the influence of different frequency components on model predictions. The proposed method enables more precise interpretability analysis at the frequency-domain level and demonstrates particular effectiveness in tasks characterized by dense frequency features. Experimental results show that the proposed method outperforms existing methods in multiple time series classification tasks, particularly in those involving high-frequency signals. It significantly improves model interpretability and captures key frequency-domain features. Compared to benchmark methods, the proposed approach more accurately identifies the critical factors influencing prediction outcomes, thus enhancing model transparency and reliability.