Deep Neural Networks （DNNs） are highly susceptible to adversarial attacks， causing security problems in time-series data classification tasks. Gradient-based attack methods can generate adversarial samples quickly but need continuous access to the model's internal information， while generation-based attack methods do not need this access after training but suffer from poor stealthiness and transferability. To address these problems， a semi-white box adversarial sample generation method for time-series data based on local augmentation was proposed using the generative attack method AdvGAN. The local augmentation strategy in this method injected information from other data categories into original samples and utilized enhanced data to execute semi-white-box attacks. The attack model leveraged both original sample information and distribution characteristics of other categories， thereby enhancing model's attack capability and transferability. Experimental results on UCR datasets demonstrate that the proposed method generates an adversarial example in 0.027 s on average； it outperforms Fast Gradient Sign Method （FGSM）， AdvGAN， and GATN （Gradient Adversarial Transformation Network） methods in attack success rate on 18， 25， and 13 datasets of 27 datasets respectively. The generated adversarial examples exhibit significantly smaller Mean Squared Error （MSE） compared to AdvGAN and GATN methods on 20 and 27 datasets respectively. Its transfer success rates surpass AdvGAN and FGSM methods on 18 and 11 datasets respectively， with transfer attack success rates exceeding 25% on 9 datasets of 21 datasets. The results indicate that the proposed method maintains efficient adversarial example generation while improving stealthiness and preserving competitive attack performance.