Sequential Recommendation Systems aim to utilize past interaction behaviors to model user interests and predict their next potential interaction item. Most existing studies model users’ interests based on complete interaction sequences. However, in practical scenarios, user interaction sequences inevitably contain inherent noise, such as incorrect clicks. The coexistence of noise and valid interactions can severely disrupt the learning of user interest representations, causing learned user interest representations diverging from the user’s actual preferences. Furthermore, some studies attempt to reduce the impact of noise by modeling the contextual information of sequences, which still have certain limitations when the contextual information itself contains noise. To tackle the above problems, we propose a Continuous Sub-sequence Denoising framework(CSD) for Sequential Recommendation. Firstly, through the dynamic partitioning and representation fusion mechanism, the long-term and short-term interests of users are accurately captured under the guidance of the target item. Then, we design an iterative denoising module that uses user long-term interests as the supervisory signal to guide the denoising process. This module identifies noise sub-sequences that deviate from users’ interest representations through a dual-judgment mechanism, enabling the precise identification of continuous noise data within the sequence. Experimental results on four large real-world datasets demonstrate that the CSD outperforms the existing denoising methods and can be effectively integrated into various mainstream sequential recommendation models.